Coal Mine Safety Alert System: Refining BP Neural Network with Genetic Algorithm Optimization

Hanzhe Pan; Jiabin Luo

REVIEW 5 major objections 8 minor 15 references

Coal Mine Safety Alert System: Refining BP Neural Network with Genetic Algorithm Optimization

T0 review · 5 major / 8 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper claims that a genetic algorithm-optimized backpropagation network (GA-BP) predicts coal-mine safety risk with lower initial error, roughly 20 percent less error, and faster convergence than a standard BP network on the same…

desk verdict A routine GA-BP application to coal mine safety that honestly reports its own limits, but the claimed advantage over plain BP is not supported by the evaluation. read the letter →

arxiv 2412.04487 v1 pith:ZVCCBDSN submitted 2024-11-21 cs.NE

Jiabin Luo , Hanzhe Pan This is my paper · ORCID

classification cs.NE

keywords coalminesafetyearlywarningsystemgeneticalgorithmbackpropagationneuralnetworkGA-BPindicatorssmall-sampleprediction

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a backpropagation neural network whose initial weights and thresholds are tuned by a genetic algorithm (GA-BP) predicts coal-mine safety risk more accurately than a plain BP network on a small real-world dataset. Using 19 safety indicators and 46 normalized samples from a Shandong mine, the authors report that GA-BP's initial error stays below 0.20 while BP's starts above that, with about a 20 percent error reduction and faster convergence. If right, the optimization scheme offers a practical upgrade path for mine early-warning systems without changing the underlying neural architecture. The paper itself notes the absolute error is still too high for operational use, which it blames on the small sample and data quality.

What carries the argument

The central object is the GA-BP hybrid: a three-layer feedforward network (19 input nodes for the safety indicators, 11 hidden nodes, 1 output node) whose weight matrix $W$ (input to hidden), hidden thresholds, weight matrix $V$ (hidden to output), and output thresholds are concatenated into one chromosome of length 232. The genetic algorithm evaluates each chromosome by running forward propagation with fitness $F = 1/(\hat{y} - y)^2$, then applies roulette-wheel selection, real-number crossover, and mutation to evolve better initial weights before BP training. This gives BP a better starting point in weight space, which is what carries the claimed improvement.

What would settle it

Run both BP and GA-BP on the same 46 samples with repeated random initializations and a leave-one-out cross-validation; if GA-BP's mean squared error is not consistently below BP's across folds, the claimed advantage is an artifact of the single 10/3 split. A simpler check is whether the stated ~20 percent error reduction reproduces when the test periods are chosen differently.

Watch

Extended reading notes

Core claim

The central claim is that encoding the BP network's four weight and threshold matrices as a real-valued chromosome and searching for good initial values with a genetic algorithm—using selection, crossover, and mutation with fitness $F = 1/(\hat{y} - y)^2$—produces a network that starts at lower error and converges faster than standard backpropagation. On the 46-sample Shandong dataset with 10 training and 3 test periods, GA-BP reports initial errors below 0.20 versus BP above 0.20, a roughly 20 percent error reduction after GA optimization, and faster approach to the target in both training and test sets. The authors claim this makes the model capable of timely warnings, while acknowledging that the overall error remains too high for satisfactory operational performance.

Load-bearing premise

The 46 normalized samples from the cited thesis are accurate and representative, and the arbitrary split of 10 training and 3 test periods captures the range of safety conditions well enough to judge the model.

Editorial extensions

If this is right

Coal mines with similar indicator systems can adopt GA-BP to get lower initial error and faster convergence than plain BP, at modest computational cost.
The approach extends to any small-sample safety early-warning problem where the indicators can be normalized to 0-1 and labeled with risk levels.
Because the GA only optimizes starting weights, it can be combined with other BP variants or training algorithms without changing the network structure.
The reported 20 percent error reduction implies that optimization of initialization matters more than architecture changes for this kind of dataset.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same GA-initialization trick would likely help other gradient-based learners on tiny datasets, but the 10/3 train-test split is too small to establish generalization; a leave-one-out cross-validation would be a stiffer test.
The high absolute error the authors concede suggests the bottleneck is data volume, not initialization; GA-BP can only partly compensate.
The indicator set (personnel behavior, equipment state, environment, management) is generic enough that GA-BP could be tested on other mines' data to see whether the 20 percent improvement is specific to this mine or a general effect.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick read for you on arXiv:2412.04487. The paper is a small case study applying a standard GA-BP hybrid to a coal-mine safety early-warning dataset taken from a 2023 thesis. It does not introduce a new algorithm or theory; the contribution is the specific application and the indicator system. To its credit, the paper is clearly organized, describes the GA-BP procedure in enough detail to reproduce in principle, and is refreshingly honest about the model's poor overall performance. That honesty is not enough to rescue the central claim, though.

The claim that GA-BP beats plain BP is not established by the evidence. The comparison rests entirely on Figure 4, with no numerical test-set errors, no error bars, no statistical tests. The test set has three samples, chosen without explanation. Worse, the min-max normalization in Eq. (11) is applied to all 46 samples before the temporal split, so the extrema of the future test period are visible during training. That is a real methodological leak: the evaluation does not simulate a forecasting setting. With n_test=3, even without the leak, random variation could produce the plotted gap. The authors' own statement that 'the overall operational outcomes remain unsatisfactory due to high error rates' further undercuts the practical significance.

The paper does cite relevant Chinese literature and gives a reasonable set of safety indicators, but the citation pattern is fine. The data and code are not available, and hyperparameters are hand-picked. The load-bearing assumption—that this one 10/3 split, normalized with future extrema, demonstrates GA-BP's superiority—is not supported.

Who is this for? A reader with a narrow interest in GA-BP applications in mining safety might glance at it, but no one should rely on the comparison. The paper is a routine engineering write-up, not a substantive research result. It deserves neither a serious referee nor a citation. If the authors had released data and code, and evaluated on a proper temporal split with multiple runs and significance testing, it could be a modest engineering note. As it stands, desk rejection is appropriate.

Referee Report

5 major / 8 minor

Summary. The manuscript proposes a hybrid genetic-algorithm-optimized backpropagation network (GA-BP) for coal-mine safety early warning. The model is a three-layer BP network with 19 input neurons (from the four-category index system in Table 1), a single output neuron representing the warning level, and 11 hidden neurons; a genetic algorithm is used to optimize the initial weights and thresholds, with chromosome length 232 per Eq. (6), fitness F = 1/E, and selection, crossover, and mutation operations described in Section 2. The case study in Section 3 uses 46 min-max-normalized samples from a Shandong coal mine, taken from the dissertation [15]; 10 samples (September 2008-June 2011) are used for training and 3 samples (December 2014-June 2015) for testing. Results are presented only as plots in Figure 4. The paper claims that GA-BP has a significant advantage over plain BP from the start (initial error below 0.20 versus above 0.20), achieves an approximate 20% error reduction, and converges faster, while also acknowledging that the overall operational outcomes remain unsatisfactory because of the small sample size and low data quality.

Significance. If the comparative claim were rigorously established, this would be a modest engineering application note: the hybrid GA-BP scheme itself is a standard technique, so the methodological novelty is limited to its application to coal-mine safety early-warning data. The topic is practically important given China's coal-mine accident record, and the paper is transparent about its parameter settings (Table 3), about the provenance of the data ([15]), and about the poor final accuracy of the models, which is a point in its favor. However, as presented, the central claim is supported only by qualitative inspection of a single figure, with no quantitative test-set metrics, no repeated stochastic trials, no error bars, and effectively three test samples, so the significance of the reported result is low even if the underlying effect is real.

major comments (5)

[§3, Fig. 4] The central claim of GA-BP superiority rests entirely on the plots in Figure 4; the paper reports no quantitative test-set error metric (MSE, MAE, accuracy, or warning-level hit rate), and the statements about initial error below 0.20 and an approximate 20% error reduction cannot be verified from the figure as printed. Because both BP training and the genetic algorithm are stochastic and the test set contains only three samples, a single run of each model could easily produce the displayed pattern by chance; repeated runs with error bars or a paired statistical test are required to support the word 'significant'.
[§3, train/test split] The evaluation uses ten training samples (September 2008-June 2011) and three test samples (December 2014-June 2015) drawn from the 46 normalized rows of Table 2, but the paper never explains why that particular split was chosen, why the sizes are 10 and 3, or what happened to the remaining 33 samples. With n = 3 test samples and a single arbitrary split, no generalization claim about early-warning capability can be supported, and this protocol is load-bearing for the abstract's claim of demonstrated superiority.
[§3, Eq. (11)] The min-max normalization of Eq. (11) is applied to the full dataset before the temporal split, so the column extrema x_j^{min} and x_j^{max} are computed over samples that include the December 2014-June 2015 test period. This makes the future test-period range visible to the training inputs and means the evaluation does not simulate a genuinely prospective forecasting setting; the stress-test concern about future-extrema leakage into training is therefore valid.
[§3, final paragraph; §4; Abstract] The manuscript contradicts its own headline result: Section 3 states that 'the overall operational outcomes remain unsatisfactory due to high error rates,' while the Abstract and Section 4 claim that results aligned with expected objectives and that the model can identify potential safety risks promptly. If the error rates are as high as the paper concedes, the reported advantage of GA-BP over BP does not establish a usable early-warning system, and the paper's own assessment undercuts the practical-significance claim made in the Abstract.
[§2, Eq. (10)] The description of the GA mutation in Section 2 says that the mutation operation 'selects the first gene of the first individual for mutation,' which, taken literally, cannot maintain population diversity as claimed; Eq. (10) is rendered too garbled to check (the two cases for a_{ij} are not typeset correctly). Because the GA-BP algorithm is exactly the method whose performance is being evaluated, this inconsistency makes the reported results irreproducible from the description.

minor comments (8)

[§2, Eq. (1)] Eq. (1) is unreadable: the hidden-layer formula is rendered as 'n mn a + = +', and the surrounding sentence about the integer range [1,10] does not explain how the value 11 follows.
[Table 3] Table 3 presents the BP and GA hyperparameters (hidden-layer node count, population size, generations, crossover and mutation rates, learning rate, performance goal) with no sensitivity analysis, so the dependence of the reported comparison on these hand-set values is unknown.
[Table 2] Table 2 replaces most of the 46 rows with ellipses, so the reader cannot reconstruct the dataset; the full normalized data should be supplied as a supplement together with a description of how the raw values from [15] were processed.
[§3] The 'approximately 20% reduction in error' statement in Section 3 is unsupported by any numerical computation or by a readable figure from which the percentage could be derived.
[Fig. 4] Figure 4 consists of low-resolution screenshots whose axes and curve values are illegible at publication scale, which matters because the paper's central claim is based on the visual pattern of these curves.
[§2, Eqs. (2)-(5)] Section 2 refers to 'four weight matrices' but Eqs. (2)-(5) define two weight matrices and two threshold matrices, and the five warning levels (high, higher, medium, lower, low) are never mapped to numeric output ranges, leaving the interpretation of the single output neuron undefined.
[§1, language] The English throughout the manuscript needs substantial revision; for example, Section 1 contains the sentence 'It occurred on February 22, 2023, the Inner Mongolia Xinjiang coal mine disaster, the death toll of 53 people, the accident since the founding of New China, China's largest accident in open-cast coal mines,' which is ungrammatical and conflates the accident date with the reporting date.
[References 12 and 14] References 12 and 14 contain stray fragments such as '495.', '496.', '500', and '501' that appear to be artifacts of database export and should be removed.

Circularity Check

1 steps flagged · score 6.0 of 10

GA-BP's headline 'initial error' advantage is the GA's own fitness objective, so the main quantitative evidence reduces to the optimization setup by construction.

fitted input called prediction [Section 2 (GA fitness function) and Section 3 (Simulation Results, Figure 4)]
"E=(y^-y)2 as the objective function, thus calculating the fitness function F=1/E. ... The Genetic Algorithm optimized Backpropagation (GA-BP) neural network exhibits a significant advantage over the traditional Backpropagation (BP) neural network right from the start, with initial error values below 0.20 compared to the BP network's values above this threshold. This initial edge is further enhanced by an approximate 20% reduction in error achieved through genetic algorithm optimization, leading to a substantial increase in accuracy."

The 'initial error' cited as proof of GA-BP's superiority is exactly the squared output error E=(y^-y)^2 that the genetic algorithm is constructed to minimize through F=1/E. Comparing the GA-optimized starting point with a randomly initialized BP's starting error is comparing the optimizer's objective value against a random guess, not an empirical prediction about generalization. The claimed '20% reduction in error achieved through genetic algorithm optimization' is likewise the GA's own objective-function improvement, so it holds by construction. The paper provides no quantitative test-set error or statistical comparison to establish that the advantage persists out of sample; the central headline evidence therefore reduces to the definition of the fitness function.

full rationale

The only genuine circular step is the paper's use of the GA's fitness value as evidence of predictive superiority. The genetic algorithm optimizes weights and thresholds by minimizing E=(y^-y)^2, and the paper then reports that GA-BP starts with lower error than BP, calling this a 'significant advantage' and a '20% reduction in error.' That comparison is an artifact of the optimizer's objective, not an independent evaluation. The paper also claims better training- and test-set performance and faster convergence, which could be empirical content, but no test-set error values, error bars, or significance tests are reported; with only three test samples and min-max normalization performed before the temporal split (Eq. 11, Table 2), the independent content is not substantiated. There is no load-bearing self-citation: the cited works are not the authors' own, and no uniqueness theorem or prior ansatz is imported. The methodological issues of tiny test size, data leakage through normalization, and hand-set hyperparameters are correctness risks rather than circularity per se, but they compound the fact that the headline quantitative claim is by construction. Overall, the central claim is partially circular: the strongest numerical evidence reduces to the GA's fitness definition, while the remaining empirical support is not quantified. Score 6 reflects this construction-level circularity, not a full equivalence between derivation and input.

Assumptions & free parameters 7 free parameters · 3 assumptions · 0 invented entities

The central claim rests on hand-chosen hyperparameters, a domain-specific indicator list, and an external dataset of unverified quality. No new physical or mathematical entities are introduced.

free parameters (7)

Hidden layer neuron count = 11
Set from a heuristic formula with an unspecified additive constant; affects the capacity of the network.
GA population size = 60
Chosen by hand in Table 3; no sensitivity analysis.
GA generations = 500
Chosen by hand in Table 3.
Crossover probability = 0.7
Chosen by hand in Table 3.
Mutation rate = 0.05
Chosen by hand in Table 3.
Learning rate = 0.001
Chosen by hand in Table 3.
Performance goal = 0.00001
Target MSE for training, chosen by hand.

assumptions (3)

domain assumption The 19 level-2 safety indicators in Table 1 fully capture the factors determining coal mine safety.
The model output is derived from these inputs; if key factors are missing, the warning levels are not trustworthy.
domain assumption The normalized dataset from literature [15] is accurate and the five-level warning labels are correct.
All training and test results inherit the quality of this external dataset.
ad hoc to paper A single hidden layer with a heuristic neuron count is sufficient for this classification task.
Equation (1) and the choice of 11 hidden neurons are heuristic; no cross-validation is shown.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Coal Mine Safety Alert System: Refining BP Neural Network with Genetic Algorithm Optimization." pith.science (2026). https://pith.science/paper/ZVCCBDSN

@misc{pith2026241204487,
  author       = {Pith},
  title        = {Pith review of: Coal Mine Safety Alert System: Refining BP Neural Network with Genetic Algorithm Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZVCCBDSN}},
  note         = {Machine review of arXiv:2412.04487}
}

read the original abstract

In response to the persistent safety challenges within coal mines, this study proposes a novel approach integrating a three-layer feedforward backpropagation artificial neural network with a genetic algorithm (GA-BP) for establishing a safety early warning system. Focused on a coal mine in Shandong, China, the model's effectiveness is evaluated using relevant data for training and analysis. Results indicate the superiority of the GA-BP model over traditional BP neural networks, offering enhanced capability for identifying potential safety risks promptly. This advancement enables coal mine management to implement timely interventions, ensuring the safety of miners. The findings present valuable insights for engineering applications in similar contexts.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 15 canonical work pages

[15]

Research on safety early warning of coal mining face based on data mining

Meng Fanqiang. Research on safety early warning of coal mining face based on data mining. University of Science and Technology Beijing,2023.DOI:10.26945/d.cnki.gbjku.2022.000369

work page arXiv 2023
[1]

Foresight. China Coal Industry Supply and Demand Market Analysis Coal Overall Output Shows Growth Trend Industry Research Report 2023.https://www.qianzhan.com/analyst/detail/220/230330308e072c.html

work page 2023
[2]

National coal mine safety production accidents in 2022.https://www.mkaq.org/html/2023/02/16/657155.shtml

Coal Mine Safety Production Network. National coal mine safety production accidents in 2022.https://www.mkaq.org/html/2023/02/16/657155.shtml

work page 2022
[3]

Accident Express

CoalMineSafety.com. Accident Express. Latest Coal Mining Accidents Coal Mining AccidentCases.CoalMineAccidentExpress.https://www.mkaq.org/sggl/shigukb/

work page
[4]

Coal and gas protrusion hierarchical-expandable early warning technology and application

GUO Deyong, HU Jie, WANG Yankai. Coal and gas protrusion hierarchical-expandable early warning technology and application. Chinese Journal of Safety Science, 2017,27(01):88-92

work page 2017
[5]

An intelligent ventilation management method for coal mines based ona rough set and improved capsulenetwork

Su Shuxian, Ouyang Meisan. An intelligent ventilation management method for coal mines based ona rough set and improved capsulenetwork. Coal Science and Technology, 2021,49(7):88-92

work page 2021
[6]

Analysis of the application of the Box-Jenkins model in coal mine roofsafetymonitoring.ElectronicDevices,2019,42(03):802-805

LI Wenfeng, BAI Hui. Analysis of the application of the Box-Jenkins model in coal mine roofsafetymonitoring.ElectronicDevices,2019,42(03):802-805

work page 2019
[7]

Chen.Design ofcoal mineworkingface safety early warningsystem basedonsafety evaluation system

H.Z. Chen.Design ofcoal mineworkingface safety early warningsystem basedonsafety evaluation system. Inner Mongolia Coal Economy,2022(22):87- 89.DOI:10.13487/j.cnki.imce.023011

work page doi:10.13487/j.cnki.imce.023011 2022

Show all 15 references

[8]

DingRijia,ZhangYibing.Early warningofmineecological safety riskbased onAHP-IE- MEAmodel.ResourcesandEnvironmentofThousandDryAreas,2017,31(10):83-89

2017
[9]

Research on safety behavior evaluation and early warning of coal mine employees.AnhuiUniversityofTechnology,2017

Yanna Zhu. Research on safety behavior evaluation and early warning of coal mine employees.AnhuiUniversityofTechnology,2017

2017
[10]

Research on the application of big data visualization predictive analysis engine in coal mine water damage monitoring and early warning

Zhao Yanchao, Li Peng, WU Tao et al. Research on the application of big data visualization predictive analysis engine in coal mine water damage monitoring and early warning. Coal and Chemical Industry,2022,45(05):53- 57+62.DOI:10.19286/j.cnki.cci.2022.05.016

2022 doi
[11]

Construction of an intelligent early warning platform based on cloud service for coal mine water damage monitoring big data

Qiao Wei, JIN Dewu, WANG Hao, et al. Construction of an intelligent early warning platform based on cloud service for coal mine water damage monitoring big data. Journal ofCoal,2020,45(07):2619-2627

2020
[12]

A comprehensive evaluation of the working environment and risk of coal and gasoutburst[dissertation].Anhui:495.AnhuiUniversityTechnology;2015.496

Guan WJ. A comprehensive evaluation of the working environment and risk of coal and gasoutburst[dissertation].Anhui:495.AnhuiUniversityTechnology;2015.496

2015
[13]

Basic research on real-time monitoring and warning technology of water inrush from coal 497 floor

Jin DW, Zheng G, Fan J. Basic research on real-time monitoring and warning technology of water inrush from coal 497 floor. In: Proceedings of the 2013 Academic Annual Conference of China Occupational Safety and Health Association; 498 2013. pp. 1029- 1036.499

2013
[14]

Study and application of comprehensive evaluation of coal mine dust hazard [dissertation].Taiyuan:Taiyuan500UniversityTechnology;2014.501

Liu X. Study and application of comprehensive evaluation of coal mine dust hazard [dissertation].Taiyuan:Taiyuan500UniversityTechnology;2014.501

2014

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [15]

Research on safety early warning of coal mining face based on data mining

Meng Fanqiang. Research on safety early warning of coal mining face based on data mining. University of Science and Technology Beijing,2023.DOI:10.26945/d.cnki.gbjku.2022.000369

work page arXiv 2023

[2] [1]

Foresight. China Coal Industry Supply and Demand Market Analysis Coal Overall Output Shows Growth Trend Industry Research Report 2023.https://www.qianzhan.com/analyst/detail/220/230330308e072c.html

work page 2023

[3] [2]

National coal mine safety production accidents in 2022.https://www.mkaq.org/html/2023/02/16/657155.shtml

Coal Mine Safety Production Network. National coal mine safety production accidents in 2022.https://www.mkaq.org/html/2023/02/16/657155.shtml

work page 2022

[4] [3]

Accident Express

CoalMineSafety.com. Accident Express. Latest Coal Mining Accidents Coal Mining AccidentCases.CoalMineAccidentExpress.https://www.mkaq.org/sggl/shigukb/

work page

[5] [4]

Coal and gas protrusion hierarchical-expandable early warning technology and application

GUO Deyong, HU Jie, WANG Yankai. Coal and gas protrusion hierarchical-expandable early warning technology and application. Chinese Journal of Safety Science, 2017,27(01):88-92

work page 2017

[6] [5]

An intelligent ventilation management method for coal mines based ona rough set and improved capsulenetwork

Su Shuxian, Ouyang Meisan. An intelligent ventilation management method for coal mines based ona rough set and improved capsulenetwork. Coal Science and Technology, 2021,49(7):88-92

work page 2021

[7] [6]

Analysis of the application of the Box-Jenkins model in coal mine roofsafetymonitoring.ElectronicDevices,2019,42(03):802-805

LI Wenfeng, BAI Hui. Analysis of the application of the Box-Jenkins model in coal mine roofsafetymonitoring.ElectronicDevices,2019,42(03):802-805

work page 2019

[8] [7]

Chen.Design ofcoal mineworkingface safety early warningsystem basedonsafety evaluation system

H.Z. Chen.Design ofcoal mineworkingface safety early warningsystem basedonsafety evaluation system. Inner Mongolia Coal Economy,2022(22):87- 89.DOI:10.13487/j.cnki.imce.023011

work page doi:10.13487/j.cnki.imce.023011 2022