Outburst danger early warning analysis method and system suitable for being applied to heading machine

Abstract

The invention discloses an outburst danger early warning analysis method and system suitable for being applied to a heading machine, and relates to the technical field of coal mine safety. The method comprises the following steps: acquiring a working face coal seam image, a personnel image, gas concentration, coal seam stress, wind speed, a ground stress impact energy peak value and coexistence rate acquired by acquisition equipment of an outburst early warning electric control platform additionally arranged on the heading machine through wireless communication; preprocessing and analyzing working face coal seam and personnel images to obtain personnel states and bearing strength of soft rock and hard rock, calculating change rates of gas concentration and wind speed, and comparing coal seam stress with the bearing strength of the soft rock and the hard rock to calculate stress intensity; according to the method, the impact energy peak value and the coexistence rate are jointly brought into consideration of an adaptive fuzzy prediction model based on collaborative decision making, data change correlation is analyzed to evaluate the outburst early warning level, the personnel state and the outburst early warning level are analyzed in combination with a mine G-IS technology, a knowledge graph and a rule engine, and an optimal emergency scheme is matched. And real-time monitoring and early warning coping of outburst dangers are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of coal mine safety, and in particular to a method and system for early warning and analysis of outburst danger suitable for application on a roadheader. Background Art

[0002] In recent years, with the widespread popularization and application of roadheaders in coal mine roadway driving, the outburst accidents induced by roadheaders in coal roadway driving have increased year by year. According to the statistics of outburst accidents occurring in the past three years, the outburst accidents induced by roadheaders in coal roadway driving have exceeded 70% of the total outburst accidents. Especially in some low-gas coal mines where outburst accidents have never occurred before, outburst accidents frequently occur due to the use of roadheaders for driving. According to the analysis of accident investigation data, the vast majority of outbursts induced by roadheaders occur during the construction process of cutting coal or erecting supports by the roadheader. At this time, the construction workers are near the working face, which often causes major and serious injury accidents. The reasons are as follows:

[0003] 1) The roadheader has a fast driving speed and a large roadway space developed, which intensifies the stress tension degree of the working face, the disturbed stress increases and is very active. The increased disturbed stress is superimposed on the residual tectonic stress encountered during roadway driving, increasing the risk of outburst triggered by in-situ stress.

[0004] 2) Using a roadheader for driving has a small impact force on the working face and surrounding rock of the roadway, making the stress concentration area in front of the working face easier to approach the working face, weakening the safety protection distance of the coal body. Without any detection equipment, it is very difficult for the operators of the roadheader to detect the stress concentration phenomenon on the working face, making it easy for in-situ stress to trigger and induce outbursts.

[0005] 3) The concentrated stress near the working face further compacts the coal seam, reducing the gas permeability of the coal seam. The gas concentration in the roadway air flow does not increase, and even is lower than the gas concentration under normal conditions. If the change and increasing degree of in-situ stress are not known, it is often considered that there is no outburst danger at the working face at this time, and the driving is accelerated, further intensifying the stress concentration degree and leading to outbursts.

[0006] 4) In the working face where a roadheader is used for driving, the operating personnel work near the heading face. Once an outburst occurs, due to the lack of early warning prompt information, accidents often occur.

[0007] Therefore, it is necessary to research and develop a prominent danger early warning analysis method and system suitable for application on roadheaders. There are still some deficiencies in the prediction and early warning methods in the existing technologies: Most of the existing prediction methods are based on empirical formulas or statistical models, which are difficult to accurately reflect the complex mechanism of coal and gas outbursts, resulting in limited prediction accuracy; Some early warning systems rely only on a single monitoring means, such as gas concentration monitoring, ignoring the monitoring of other key parameters, and it is difficult to comprehensively reflect the precursor information of coal and gas outbursts; Some early warning systems have the problem of lagging data update, unable to reflect the dynamic change process of coal and gas outbursts, resulting in poor early warning effects.

[0008] As disclosed in the Chinese patent with the authorization announcement number CN110118103B, a coal mine gas early warning method includes: First, obtain the gas sensor information table, gas data information table, and gas early warning data information table with configured gas early warning parameters from the gas monitoring system, and then conduct comparative analysis through the gas data information table, the gas sensor information table with configured gas early warning parameters, and the gas early warning data information table to obtain the real-time gas early warning data information table and the gas historical alarm information table. This technical solution solves the defect that the existing coal mine gas monitoring system is prone to cause accidents by alarming only after gas exceeds the limit.

[0009] The above existing technologies all have the following problems: Lack of prominent early warning ability and spatial positioning ability, unable to detect potential dangers in advance and take emergency measures. Summary of the Invention

[0010] In view of the deficiencies in the existing technologies, the present invention proposes a prominent danger early warning analysis method and system suitable for application on roadheaders to achieve the linkage execution of prominent early warning and emergency decision-making, and improve the safety of mine operations and the efficiency of emergency response.

[0011] The technical solution adopted to achieve the object of the present invention is as follows:

[0012] A prominent danger early warning analysis method suitable for application on roadheaders includes the following specific steps:

[0013] Wirelessly receive the image I(t) and environmental information X(t) collected by the acquisition device of the prominent danger early warning electronic control platform at time t. The image I(t) includes the coal seam image Is(t) and the personnel image Ip(t) of the working face, and the environmental information X(t) includes the gas concentration x c (t), coal seam stress x s (t), wind speed x w (t), peak impact energy x e (t), and inherent rate x j (t);

[0014] The image I(t) is subjected to light source weakening and multi-transmission fusion by using the multiple transmission enhancement method, and combined with the atmospheric scattering model to generate the clear image J(t) at time t. The clear image J(t) includes the clear working face coal seam image J s (t) and the clear personnel image J p (t);

[0015] The pose characteristics of the staff in the clear personnel image J p (t) are captured by the target state recognition network to generate the personnel state z(t) at time t. The soft rock and hard rock in the clear working face coal seam image J s (t) at time t are accurately analyzed and identified by the coal seam rock mass recognition network, and the bearing strength of the soft rock is determined and the bearing strength of the hard rock

[0017] The change rate of gas concentration at time t is calculated The stress intensity gx s (t), the change rate of wind speed The change rate of gas concentration is given by using the adaptive fuzzy prediction model based on collaborative decision-making The stress intensity gx s (t), the change rate of wind speed The peak impact energy x e (t) and the coincidence rate x j (t) corresponding input factor weights are combined to construct the input factor sequence The input factor sequence is fuzzified And reasoning is carried out according to the fuzzy rules considering the change correlation of the input factors, and defuzzification is performed to generate the outburst warning level y(t) at time t;

[0018] Combined with the mine GIS technology, the position of the roadheader is obtained. Based on the knowledge graph and the rule engine, the personnel state z(t) and the outburst warning level y(t) at time t are comprehensively analyzed, and the optimal emergency plan is automatically matched and executed. The clear image J(t), the environmental information X(t), the outburst warning level y(t), the position of the roadheader and the optimal emergency plan are stored in the knowledge graph to achieve automatic update.

[0019] Specifically, the steps of generating the clear image J(t) by using the multiple transmission enhancement method to process the image I(t) are as follows:

[0020] The image I(t) is converted into an HSV image, and the luminance map I V (t) on the luminance channel V is subjected to minimum filtering, and the luminance I V (t,x) of any point x in the image I(t) is changed to the minimum luminance of all points in the local neighborhood Ω x centered at point x, and the weakened luminance map I′ V(t);

[0021] Obtain the weakened luminance image I′ V Take the maximum luminance value in (t) as the atmospheric light value A(t) at time t;

[0022] According to the dark channel principle, the n - order darkness value of any point x in the image I(t) is changed to the minimum color component of all points in the RGB three - color channels in the local block centered at point x with a scale of n, and obtain the 1 - order dark image to the N - order dark image n = 1, …, N, where N is the total number of scales;

[0023] There is a linear relationship between the ratio of the dark image and the atmospheric light value and the transmittance. Deduce the transmittance of the 1 - order dark image to the N - order dark image and take the average value as the transmittance h(t) of the image I(t);

[0024] Combined with the atmospheric scattering model, reverse - clear the image I(t) through the atmospheric light value A(t) and the transmittance h(t) to generate the clear image J(t).

[0025] Specifically, classify to generate the clear personnel image J p (t) of the personnel state z(t) includes the following steps:

[0026] Deeply extract the clear personnel image J p (t) through 2 - time normalization convolution to generate the 1 - level image information and the 2 - level image information;

[0027] Use the residual structure to superimpose and fuse the 1 - level image information and the 2 - level image information to generate multi - scale image information;

[0028] Through the spatial attention mechanism, screen the personnel spatial features in the multi - scale image information that are beneficial to the determination of the personnel pose and strengthen the correlation between the personnel spatial features and the spatial pose to obtain the personnel spatial feature map;

[0029] Through the channel attention mechanism, focus on the pixel points of the multi - scale image information from the RGB three channels respectively to generate the personnel semantic feature map that is beneficial to personnel grasping;

[0030] Superimpose the personnel spatial feature map and the personnel semantic feature map as the real part and the imaginary part respectively, and classify through the complex - number linear layer and the complex - number support vector machine to generate the personnel state z(t), where the personnel state z(t) includes actionable and non - actionable.

[0031] Specifically, accurately analyze the clear working - face coal - seam image J at time t through the coal - seam rock - mass recognition networks (t) to identify soft rock and hard rock and determine the bearing strength of soft rock and the bearing strength of hard rock including the following specific steps:

[0032] Perform depth convolution on the clear working face coal seam image J s (t) on the RGB three channels respectively to extract the spatial features of the three channels, and generate the working face feature map by combining the depth convolution results of the three channels through pointwise convolution;

[0033] Generate multiple extended feature maps by applying atrous convolution with multiple different dilation rates to the working face feature map, and then generate the aggregated extended feature map by applying pointwise convolution again;

[0034] Generate the coal-rock fusion feature map by fusing high-dimensional semantic features and low-dimensional semantic features through bilinear interpolation, 4x upsampling and feature compression in sequence for the aggregated extended feature map;

[0035] Input the coal-rock fusion feature map into 2 Softmax logistic regression models respectively for linear dimensionality reduction and non-linear activation to generate the soft rock type probability vector and the hard rock type probability vector;

[0036] Determine the bearing strength of soft rock and the bearing strength of hard rock

[0037] Specifically, the change rate of gas concentration at time t is the change amount obtained by subtracting the gas concentration x c (t) at the previous time t - Δt from the gas concentration x c (t - Δt) and then dividing by the time interval Δt;

[0038] When the coal seam stress x s (t) at time t is less than or equal to the bearing strength of soft rock , the stress intensity gx s (t) is always 0;

[0039] When the coal seam stress x s (t) at time t is greater than the bearing strength of soft rock , the stress intensity gx s (t) is the difference between the coal seam stress x s (t) at time t and the bearing strength of soft rock divided by the difference between the bearing strength of hard rock and the bearing strength of soft rock ;

[0040] Wind speed change rate The wind speed at time t is x w (t) minus the wind speed x at the previous time t - Δt w The resulting change is divided by the interval duration Δt.

[0041] Specifically, an adaptive fuzzy prediction model based on collaborative decision-making is combined with the gas concentration change rate The stress intensity gx s (t), the wind speed change rate The peak impact energy x e (t) and the inherent rate x j (t) to generate the outburst warning level y(t), which includes the following specific steps:

[0042] Multiply the gas concentration change rate The stress intensity gx s (t), the wind speed change rate The peak impact energy x e (t) and the inherent rate x j (t) by the corresponding input factor weights to generate an input factor sequence Is the i-th input factor at time t, where i = 1, 2, 3, 4, 5;

[0043] Use the Gaussian membership function to calculate the membership degrees of the i-th input factor belonging to K fuzzy sets respectively, and generate the i-th membership degree vector μ i (t), where K is the total number of fuzzy sets, and each fuzzy set corresponds to a Gaussian membership function;

[0044] Multiply the membership degrees of the first input factor To the fifth input factor Belonging to the k-th fuzzy set as the strength g of the k-th fuzzy rule at time t k (t), where k = 1, …, K, obtain the strengths of the K fuzzy rules at time t and normalize them to generate the corresponding strength weights, where the fuzzy rules and the fuzzy sets correspond one by one;

[0045] According to the conclusion parameter group α of the k-th fuzzy rule k Linearly sum all the input factors at time t to generate the corresponding total input factor f k (t), and perform a weighted sum of the corresponding total input factors according to the strength weights of the K fuzzy rules to generate the outburst warning level y(t) at time t.

[0046] Furthermore, the pre-training of the adaptive fuzzy prediction model based on collaborative decision-making includes the following specific steps:

[0047] Considering domain experience, calculate the gas concentration change rate in existing coal and gas outburst prediction research The stress intensity gxs , rate of change of wind speed peak impact energy x e and coincidence rate x j probability of being selected;

[0048] Considering mathematical analysis, obtain a historical data set, and calculate separately based on the rate of change of gas concentration stress intensity gx s , rate of change of wind speed peak impact energy x e and coincidence rate x j the first Gini value gn when classifying the historical data set 1 , the second Gini value gn 2 , the third Gini value gn 3 , the fourth Gini value gn 4 and the fifth Gini value gn 5 ;

[0049] Initialize the M-group domain experience ratio β area and the mathematical analysis ratio β math and process the historical data set separately to generate M different historical transformation data sets, where M is the total number of ratio groups; the domain experience ratio is the first ratio, and the mathematical analysis ratio is the second ratio;

[0050] Divide each historical transformation data set into a training set and a test set, and determine the K fuzzy rules of each historical transformation data set and the conclusion parameter group corresponding to each fuzzy rule;

[0051] Select and fix the input factor weight sequence, K fuzzy rules and the conclusion parameter group corresponding to each fuzzy rule of the optimal historical transformation data set, and the pre-training is completed.

[0052] Furthermore, determining all fuzzy rules and the corresponding conclusion parameter groups corresponding to the collaborative decision-making based adaptive fuzzy prediction model based on the historical transformation data set includes the following specific steps:

[0053] Obtain the training set and the test set, generate 1 fuzzy rule for the first input historical transformation sample in the training set, and use the variance of the historical transformation samples in the training set and the first input historical transformation sample as the initial width and the initial center to generate the first fuzzy set;

[0054] Input the second historical transformation sample in the training set, fuzzify it according to the existing Gaussian membership function and calculate the corresponding intensity g 1 , and compare it with the coverage threshold;

[0055] If it is less than the coverage threshold, generate the second fuzzy rule, remove the first historical transformation sample from the training set, recalculate the variance, and use it as the initial width of the second Gaussian membership function. Use the historical transformation sample of the second input as the initial center of the second Gaussian membership function. Generate the second fuzzy set and continue to input the third historical transformation sample in the training set.

[0056] If it is greater than or equal to the set coverage threshold, continue to input the third historical transformation sample in the training set.

[0057] Repeat the above logic until the entire training set is traversed, and a total of K fuzzy rules, K Gaussian membership functions, and K fuzzy sets are obtained. Among them, the fuzzy rules, Gaussian membership functions, and fuzzy sets correspond one by one.

[0058] Use the least squares method to fit the linear combination form of the total input factors f corresponding to the K fuzzy rules 1 ,…,f K to solve the initial conclusion parameter groups corresponding to the K fuzzy rules. Among them, and are the initial conclusion parameter groups corresponding to the first fuzzy rule and the Kth fuzzy rule respectively.

[0059] Verify the rules through the test set and calculate the F1-Score of the test set. Use the adaptive optimization algorithm to adjust the widths, centers of the K Gaussian membership functions, and the conclusion parameter groups corresponding to the K fuzzy rules to determine the widths, centers of the K Gaussian membership functions and the conclusion parameter groups corresponding to the K fuzzy rules of the historical transformation data set.

[0060] Specifically, the knowledge graph needs to be pre-constructed and can be continuously updated. The construction of the knowledge graph includes the following specific steps:

[0061] Clarify the application scenarios of the knowledge graph as mine safety monitoring, accident prevention, and emergency response, and define the schema layer of the knowledge graph.

[0062] Pre-collect mine safety data and perform cleaning, denoising, and formatting processing.

[0063] Use NLP technology for text tokenization, part-of-speech tagging, named entity recognition, and relationship extraction to identify mine safety entities, attributes, and relationships in the text.

[0064] Apply image recognition methods to identify mine safety elements in images and extract text information in the images.

[0065] Fuse the knowledge extracted from text and image knowledge, solve the entity alignment and relationship conflict problems, and store it in the graph database in the form of Neo4j.

[0066] Specifically, the rule engine analyzes the personnel status z(t) and the prominent warning level y(t) based on the conditional judgment logic, and automatically formulates the optimal emergency plan by invoking the knowledge of the knowledge graph. The conditional judgment logic includes:

[0067] If the personnel status z(t) is actionable and the prominent warning level y(t) = 0, the matched optimal emergency plan is empty;

[0068] If the personnel status z(t) is actionable and the prominent warning level y(t) = 1, the matched optimal emergency plan is to continue working intermittently and send it to the staff;

[0069] If the personnel status z(t) is actionable and the prominent warning level y(t) = 2, the matched optimal emergency plan is to stop work, combine the mine GIS technology and the heuristic algorithm to determine the optimal evacuation route from the personnel location to the mine exit and feedback it to the staff;

[0070] If the personnel status z(t) is non-actionable, the matched optimal emergency plan is to stop work and dispatch rescue, combine the mine GIS technology and the heuristic algorithm to determine the optimal rescue route from the mine exit to the personnel location and send it to the rescue personnel.

[0071] Furthermore, combining the mine GIS technology and the heuristic algorithm to determine the optimal evacuation route or the optimal rescue route includes the following specific steps:

[0072] Based on the mine GIS technology, restore the three-dimensional map of the mine and divide it into multiple regional grids by using a three-dimensional grid. Eliminate the regional grids corresponding to the prominent warning level y(t) = 2 and attach a danger mark to the neighboring regional grids to determine the starting regional grid and the target regional grid;

[0073] Select the A-Star algorithm for path planning, establish a parent list and a child list, and place the starting regional grid into the parent list;

[0074] Starting from the starting regional grid, at each step of the A-Star algorithm, calculate the evaluation value of each neighboring regional grid in the child list, and select the neighboring regional grid with the highest evaluation value to place it in the parent list. Among them, the evaluation value is equal to the Manhattan distance from the neighboring regional grid to the target regional grid minus the penalty term;

[0075] Judge whether the neighboring regional grid is the target regional grid. If it is not the target regional grid, use the neighboring regional grid as the current regional grid and continue to determine the next neighboring regional grid to go to;

[0076] If it is a target area cell, generate the optimal evacuation route or the optimal rescue route according to the storage order of the area cells in the parent list.

[0077] A prominent danger early warning analysis system suitable for application on a roadheader, used to implement a prominent danger early warning analysis method suitable for application on a roadheader, including a wireless communication module, a comprehensive early warning module, an emergency response module, and a storage module;

[0078] The wireless communication module periodically receives the image I and the environmental information X collected by the prominent warning electronic control platform and wirelessly transmitted at an interval of Δt;

[0079] The comprehensive early warning module preprocesses the image I using the multiple transmission enhancement method to generate a clear image J, and obtains the personnel status z, the soft rock bearing strength and the hard rock bearing strength Calculate the gas concentration change rate The stress intensity gx s and the wind speed change rate Use an adaptive fuzzy prediction model based on collaborative decision-making for fuzzy inference to generate a prominent warning level y;

[0080] The emergency response module combines the mine GIS technology to obtain the position of the roadheader, comprehensively analyzes the personnel status z and the prominent warning level y based on the knowledge graph and the rule engine, and matches the optimal emergency plan;

[0081] The storage module stores the knowledge graph and updates it dynamically.

[0082] Furthermore, the comprehensive early warning module includes a data processing unit, an image analysis unit, and a risk assessment unit;

[0083] The data processing unit uses the multiple transmission enhancement method to weaken the light source and perform multi-transmission fusion on the image I, and combines the atmospheric scattering model to generate a clear image J;

[0084] The image analysis unit analyzes the clear personnel image J through the target status recognition network p to generate the personnel status z, and identifies the support rock mass type in the clear working face coal seam image J through the coal seam rock mass recognition network s and determines the soft rock bearing strength and the hard rock bearing strength

[0085] The risk assessment unit calculates the gas concentration change rate The stress intensity gx s and the wind speed change rate Use an adaptive fuzzy prediction model based on collaborative decision-making to consider the gas concentration change rate The stress intensity gx s, wind speed change rate Peak impact energy x e and the co - ownership rate x j The change correlation is used to generate a prominent early warning level y through fuzzy inference and defuzzification.

[0086] Compared with the prior art, the present invention receives real - time images and environmental information through a wireless transmission method, pre - processes the images through a multiple - transmission enhancement method to solve the problem of image blurring caused by dim environments, assists the target state recognition network and the coal seam rock mass recognition network to obtain accurate personnel states and determine the bearing strength of soft rock and hard rock, calculates the change rate of gas concentration, stress intensity, and wind speed change rate, converts them together with the peak impact energy and co - ownership rate into input factors and inputs them into an adaptive fuzzy prediction model based on collaborative decision - making. According to the fuzzy rules, the change correlation of the input factors is considered for reasoning, and defuzzification is performed to generate a prominent early warning level. Combining the mine GIS technology and comprehensive analysis of personnel states and prominent early warning levels based on knowledge graphs and rule engines, the optimal emergency plan is automatically matched and executed, and the knowledge graph is updated, realizing real - time monitoring of prominent hazards at the roadheader working face and rapid execution of emergency plans, improving the safety of mine operations and the efficiency of emergency response. Brief Description of the Drawings

[0087] Figure 1 It is a flow chart of a prominent hazard early warning analysis method suitable for application on a roadheader in the present invention;

[0088] Figure 2 It is a flow chart of an adaptive fuzzy prediction model based on collaborative decision - making in the present invention;

[0089] Figure 3 It is a flow chart of determining the optimal evacuation route or the optimal rescue route by combining the mine GIS technology and a heuristic algorithm in the present invention;

[0090] Figure 4 It is a schematic diagram of a prominent hazard early warning analysis system suitable for application on a roadheader in the present invention;

[0091] Figure 5 It is a structural diagram of an outburst - prevention roadheader with the function of warning coal and gas outburst hazards in the present invention.

[0092] Reference Signs: 100, roadheader assembly; 101, roadheader body; 200, early warning assembly. Detailed Description of the Embodiment

[0093] The following further details the present invention in conjunction with the drawings and embodiments.

[0094] Embodiment 1

[0095] As Figure 1As shown in the figure, a specific embodiment of the present invention discloses a prominent danger early warning analysis method suitable for application on a roadheader, including the following specific steps:

[0096] Receive the image I(t) and environmental information X(t) collected at time t sent by the prominent warning electronic control platform on the roadheader through wireless transmission means. The image I(t) includes the coal seam image I s (t) and the personnel image I p (t) of the working face. The environmental information X(t) includes the gas concentration x c (t), the coal seam stress x s (t), the wind speed x w (t), the peak impact energy x e (t) and the co-occurrence rate x j (t). Among them, the peak impact energy x e (t) is the maximum energy value in the microseismic signals monitored by all microseismic sensors at time t. The co-occurrence rate x j (t) is the proportion of microseismic sensors that monitor microseismic signals at time t to the total number of microseismic sensors. Since the microseismic sensors are installed at different positions on the roadheader, the higher the co-occurrence rate x j (t), the greater the in-situ stress energy and range of the vibration source that generates microseismic signals at time t;

[0097] Adopt the multiple transmission enhancement method to extract the bright channel of the image I(t) and weaken the light source. Based on the dark channel principle, perform multi-transmission fusion, and combine with the atmospheric scattering model to generate the clear image J(t) at time t, effectively solving the problem of dim environment and blurred image I(t) in the coal mine underground. The clear image J(t) includes the clear coal seam image J s (t) and the clear personnel image J p (t);

[0098] Capture and analyze the pose characteristics of the staff in the clear personnel image J p (t) through the target state recognition network, and assist in classifying to generate the personnel state z(t) at time t. Accurately analyze the rock color, texture, and the contact relationship between the coal seam and the roof and floor in the clear coal seam image J s (t) at time t through the coal seam rock mass recognition network, identify soft rock and hard rock, and determine the bearing strength of soft rock and the bearing strength of hard rock Among them, the target state recognition network and the coal seam rock mass recognition network need to be pre-trained based on the personnel image and the coal seam image with additional tags. The bearing strength of soft rock and the bearing strength of hard rock can be determined by pre-collecting different types of rock masses and conducting mechanical failure tests;

[0099] Calculate the change rate of gas concentration at time t Stress intensity gx s (t), and the change rate of wind speed Adopt an adaptive fuzzy prediction model based on collaborative decision-making to obtain the change rate of gas concentration Stress intensity gx s (t), and the change rate of wind speed Peak impact energy x e (t) and the inherent rate x j (t) are respectively assigned corresponding input factor weights and combined to construct an input factor sequence Fuzzify the input factor sequence And perform reasoning according to the pre-determined fuzzy rules, and generate the outburst warning level y(t) at time t through defuzzification;

[0100] Combine the mine GIS technology to obtain the position of the roadheader, comprehensively analyze the personnel status z(t) and the outburst warning level y(t) at time t based on the knowledge graph and the rule engine, automatically match the optimal emergency plan and execute it, and store the clear image J(t), environmental information X(t), outburst warning level y(t), roadheader position and optimal emergency plan in the knowledge graph to realize the update of the knowledge graph.

[0101] Specifically, the steps of generating the clear image J(t) by using the multiple transmission enhancement method for processing the image I(t) are as follows:

[0102] Convert the image I(t) into an HSV image. Different from the conventional RGB channels of the image, H, S, and V respectively represent the hue channel, saturation channel, and brightness channel, which can intuitively express the hue, vividness, and light and dark degree of the color. Perform minimum filtering on the brightness map I V (t) of the image I(t) on the brightness channel V to weaken the influence of the light source on the image I(t). The minimum filtering changes the brightness I V (t,x) of any point x in the image I(t) to the minimum brightness of all points in the local neighborhood Ω x centered at point x. The specific formula for minimum filtering is as follows:

[0103]

[0104] where y represents any point in the local neighborhood Ω x , I′ V (t,x) is the corrected brightness of point x, I V (t,y) is the brightness of point y. After minimum filtering, the weakened brightness map I′ V (t) of the image I(t) on the brightness channel V is obtained;

[0105] Take the weakened brightness map I′ VThe maximum brightness value in I(t) is used as the atmospheric light value A(t) at time t, and the specific formula is as follows:

[0106] A(t) = I V (argmax(I′ V (t)));

[0107] Among them, I V (argmax(I′ V (t)) represents the brightness corresponding to the point with the maximum brightness in the weakened brightness map I′ V (t), that is, the maximum brightness value in the weakened brightness map I′ V (t);

[0108] According to the dark channel principle, the image I(t) is converted into dark images corresponding to N scales, which are respectively recorded as the 1st-order dark image ……, the nth-order dark image ……, the Nth-order dark image N is the total number of scales. The dark channel principle changes the value of the nth-order darkness of any point x in the image I(t) to the minimum color component of all points in the local block centered at point x and with a scale of n in the RGB three color channels, n = 1, …, N. The specific formula of the dark channel principle is as follows:

[0109]

[0110] Among them, y σ represents any point in the local block with a scale of n. The scale n determines the size of the local block. I c (t, y) represents the c color component of point x in the c color channel, c ∈ {R, G, B};

[0111] Calculate the transmission rates corresponding to the 1st-order dark image to the Nth-order dark image respectively, and record them as h 1 (t), …, h n (t), …, h N (t). Among them, h n (t) is the nth-order transmission rate corresponding to the nth-order dark image and has a linear relationship with the ratio of the nth-order dark image and the atmospheric light value A(t). The specific calculation formula is as follows:

[0112]

[0113] Among them, λ is the adjustment coefficient, A(t) is the atmospheric light value, and the 1st-order transmission rate h 1(t) to the N - order transmittance h N The mean value of (t) is taken as the transmittance of the image I(t). In this embodiment, the adjustment coefficient λ is taken as 0.85;

[0114] Combined with the atmospheric scattering model, given the known image I(t), the image I(t) can be dehazed by the atmospheric light value A(t) and the transmittance h(t), and a clear image J(t)=[I(t)-A(t)(1 - h(t))]h(t) is generated.

[0115] Specifically, a clear personnel image J is generated by classifying through the target state recognition network p The personnel state z(t) in (t) includes the following steps:

[0116] The clear personnel image J(t) is continuously and deeply extracted through two consecutive normalization convolutions, generating first - level image information and second - level image information. Among them, the normalization convolution introduces batch normalization processing on the traditional convolution to prevent gradient explosion. The first - level image information is obtained by extracting the clear personnel image through the first normalization convolution, and the second - level image information is obtained by further mining the first - level image information through the second normalization convolution; p Utilizing the skip - connection feature of the residual structure, the first - level image information and the second - level image information are superimposed and fused to generate multi - scale image information, further improving the extraction effect;

[0117] The multi - scale image information is processed through a spatial attention mechanism. The spatial attention mechanism is used to screen out the personnel spatial features in the multi - scale image information that are beneficial to the determination of the personnel pose and strengthen the correlation between the personnel spatial features and the spatial pose, obtaining a personnel spatial feature map;

[0118] The multi - scale image information is processed through a channel attention mechanism, focusing on the pixel points of the multi - scale image information from the RGB three channels respectively, attaching higher confidence to the pixel points where the personnel are located, and generating a personnel semantic feature map that is beneficial to personnel grasping;

[0119] The personnel spatial feature map and the personnel semantic feature map are superimposed as the real part and the imaginary part respectively, and a complex feature vector is generated through dimensionality reduction by a complex linear layer. Further, a complex support vector machine is used for classification to generate the personnel state z(t), where the personnel state z(t) includes actionable and non - actionable. The complex linear layer and the complex support vector machine are existing processing methods.

[0120] Specifically, the clear working - face coal - seam image J(t) at time t is accurately analyzed through a coal - seam rock - mass recognition network

[0121] to identify soft rock and hard rock and determine the bearing strength of soft rock s and the bearing strength of hard rock and the bearing strength of hard rock It includes the following specific steps:

[0122] Apply a 3×3 convolutional kernel to the clear working face coal seam image J s (t) on the RGB three channels respectively for synchronous and independent depth convolution to extract the spatial features of each channel, and perform pointwise convolution through a 1×1 convolutional kernel, and combine the depth convolution results of the three channels to generate a working face feature map;

[0123] Generate multiple extended feature maps by applying atrous convolutions with multiple different dilation rates to the working face feature map, and further perform pointwise convolution through a 1×1 convolutional kernel, and combine the multiple extended feature maps to generate an aggregated extended feature map. Among them, the atrous convolution is achieved by inserting 0s at intervals in the traditional convolutional kernel, which expands the receptive field of the convolutional kernel without increasing the amount of computation and the number of parameters, and helps the network capture more extensive context information. The dilation rate determines the number of inserted 0s;

[0124] Restore the aggregated extended feature map to the same size as the clear working face coal seam image J s (t) through bilinear interpolation, and achieve the fusion of high-dimensional semantic features and low-dimensional semantic features through 4-fold upsampling and 3×3 convolutional feature compression, and generate a coal-rock fusion feature map with the same size as the clear working face coal seam image J s (t);

[0125] Input the coal-rock fusion feature map into 2 Softmax logistic regression models respectively. The 2 Softmax logistic regression models reduce the dimension through two consecutive linear layers and use the Softmax function for non-linear activation, and generate a soft rock type probability vector and a hard rock probability type vector respectively. The dimensions of the soft rock type probability vector and the hard rock probability vector are the same as the total number of soft rock types and the total number of hard rock types respectively;

[0126] Select the soft rock type and hard rock type with the maximum probability in the soft rock type probability vector and the hard rock type probability vector, and determine the soft rock bearing strength and the hard rock bearing strength Soft rock bearing strength and the hard rock bearing strength respectively determine the in-situ stress intensities that the soft rock and hard rock can withstand.

[0127] Specifically, the change rate of gas concentration at time t is the change amount obtained by subtracting the gas concentration x c (t) at the previous time t - Δt from the gas concentration x c (t - Δt) at time t, and then dividing the change amount by the interval duration Δt. The change rate of gas concentration Reflects the gas change from the previous moment t-Δt to the moment t, and the stress intensity gx s (t) is the coal seam stress x at the moment t s (t) is less than or equal to the bearing strength of soft rock is always 0, and the working face is in a stress balance state at this time. When the coal seam stress x at the moment t s (t) is greater than the bearing strength of soft rock , the stress intensity gx s (t) is the difference between the coal seam stress x at the moment t s (t) and the bearing strength of soft rock divided by the difference between the bearing strength of hard rock and the bearing strength of soft rock , that is, the stress intensity gx s (t)=1 indicates that the coal seam stress x s (t) is equal to the bearing strength of hard rock When the stress intensity gx s (t)>1, the coal seam is impacted by the instability energy when the hard roof fractures, and the peak value of the impact energy x e (t) will mutate, which can effectively reflect whether there is an outburst risk in the working face. The wind speed change rate is the change amount obtained by subtracting the wind speed x at the previous moment t-Δt w (t) from the wind speed x at the moment t w (t-Δt) and then dividing by the interval duration Δt. The wind speed change rate reflects the wind speed change from the previous moment t-Δt to the moment t. Combining with the gas concentration change rate can effectively judge whether it is the change of the gas concentration x w (t) caused by the change of the wind speed x c (t). If the wind speed change rate is equal to 0 but the gas concentration change rate is large, combining with the stress excess amount Δx s (t) and the peak value of the impact energy x e (t) can effectively judge the outburst risk of the working face.

[0128] Such as Figure 2 shown, specifically, an adaptive fuzzy prediction model based on collaborative decision-making is used to combine the gas concentration change rate The stress intensity gx s (t), the wind speed change rate The peak value of the impact energy x e (t) and the inherent rate x j (t) to infer and generate the outburst warning level y(t), including the following specific steps:

[0129] Take the gas concentration change rate Stress intensity gx s (t), rate of change of wind speed Peak impact energy x e (t) and occupancy rate x j (t) are respectively assigned weights of input factors to generate an input factor sequence Among them, is the i-th input factor at time t, i = 1, 2, 3, 4, 5, and the specific formula is as follows:

[0130]

[0131] Among them, w i is the i-th input factor weight in the optimal input factor weight sequence determined in the pre-training stage;

[0132] The input factor sequence is input into the membership layer. The membership layer uses Gaussian membership functions to calculate the membership degrees of the i-th input factor belonging to K fuzzy sets respectively, and generates the i-th membership vector represents the membership degree of the i-th input factor belonging to the k-th fuzzy set, and the specific formula is as follows:

[0133]

[0134] Among them, o k and σ k are respectively the center and width of the k-th Gaussian membership function, k = 1, …, K, where K is the total number of fuzzy sets. Among them, the k-th Gaussian membership function corresponds to the k-th fuzzy set;

[0135] Calculate the strength g k (t) of the k-th fuzzy rule at time t. The strength g k (t) of the k-th fuzzy rule is equal to the product of the membership degrees of the 1st input factor to the 5th input factor belonging to the k-th fuzzy set, and the specific formula is as follows:

[0136]

[0137] Respectively obtain the strengths g 1 (t), …, g k (t), …, g K (t) of K fuzzy rules at time t and perform normalization processing to generate the strength weights corresponding to K fuzzy rules Among them, the fuzzy rules and the fuzzy sets are in one-to-one correspondence;

[0138] According to the conclusion parameter group α determined in the pre-training stage for the k-th fuzzy rule kLinearly sum all input factors at time t to generate the total input factor f corresponding to the k-th fuzzy rule k (t), based on the intensity weight Perform weighted summation on the total input factors f 1 (t), …, f k (t), …, f K (t) corresponding to K fuzzy rules to generate the prominent warning level y(t) at time t. The specific calculation formula is as follows:

[0139]

[0140] Where and are the first conclusion parameter and the (i + 1)-th conclusion parameter in the conclusion parameter group α k respectively. The value of the prominent warning level y(t) is defined in the pre-training stage, and y(t) ∈ {0, 1, 2}. Therefore, the collaborative decision-making based adaptive fuzzy prediction model is essentially a classification model, which classifies according to the gas concentration change rate the stress intensity gx s (t), the wind speed change rate the peak impact energy x e (t) and the inherent rate x j (t) to the corresponding prominent warning level y(t).

[0141] Furthermore, the pre-training of the collaborative decision-making based adaptive fuzzy prediction model includes the following specific steps:

[0142] Considering domain experience, which represents the degree of importance of a specific feature in a specific domain. By searching on the network in advance for the feature sets used in coal and gas outburst prediction research, calculate the selection probabilities of the gas concentration change rate the stress intensity gx s , the wind speed change rate the peak impact energy x e and the inherent rate x j respectively. Among them, the selection probability of a single feature is equal to the number of studies in which the single feature is selected divided by the total number of studies;

[0143] Considering mathematical analysis, since the gas concentration change rate the stress intensity gx s , the wind speed change rate the peak impact energy x e and the inherent rate x jFinally, it is applied to the classification task. Therefore, the Gini value is selected as the reference index. The Gini value represents the purity of the sub-dataset obtained by classifying using a single feature. The higher the Gini value, the better the effect of the single feature on classification. Obtain the historical dataset constructed based on multiple groups of pre-collected historical samples. Each group of historical samples includes the gas concentration change rate Stress intensity gx s , wind speed change rate Peak impact energy x e and the inherent rate x j and the corresponding outburst warning level. Calculate the first Gini value gn Stress intensity gx s , wind speed change rate Peak impact energy x e and the inherent rate x j when classifying the historical dataset, the second Gini value gn 1 , the third Gini value gn 2 , the fourth Gini value gn 3 , the fifth Gini value gn 4 and the fifth Gini value gn 5 . Take the first Gini value gn 1 as an example. The first Gini value gn 1 is the maximum Gini value among each value of the gas concentration change rate . The specific calculation formula is as follows:

[0144]

[0145] Among them, Num and Num 1 (x) represent the total number of historical samples and the number of historical samples with the gas concentration change rate taking the value of x respectively, is the value set of the gas concentration change rate ;

[0146] Initialize M groups of domain experience ratios β area and mathematical analysis ratios β math , and each group of domain experience ratios β area and mathematical analysis ratios β math satisfy the sum of 1. The domain experience ratio β area and mathematical analysis ratio β math determine the importance in the problem of coal and gas outburst prediction, generate M groups of input factor weight sequences. Take the first weight in the mth group as an example. The first weight in the mth group is equal to the selection probability p of the gas concentration change rate c and the domain experience ratio of the mth group The product added to the first Gini value gn 1 And the mathematical analysis ratio of the m-th group The product, for the rate of change of gas concentration in each group of historical samples Stress intensity gx s , Rate of change of wind speed Peak value of impact energy x e And the co-ownership rate x j Perform corresponding transformations to obtain M groups of input factor sequences The historical data set is processed into M different historical transformation data sets. The m-th historical transformation data set is the historical data set passing through the domain experience ratio of the m-th group And the mathematical analysis ratio Processed and generated, m = 1, …, M, where M is the total number of groups of ratios;

[0147] Divide each historical transformation data set into a training set and a test set in the same way, and determine the K fuzzy rules corresponding to the adaptive fuzzy prediction model based on collaborative decision-making and the conclusion parameter group corresponding to each fuzzy rule according to each historical transformation data set;

[0148] Select the optimal historical transformation data set with the best evaluation index among the M historical transformation data sets, obtain and fix the input factor weight sequence, K fuzzy rules and the conclusion parameter group corresponding to each fuzzy rule corresponding to the optimal historical transformation data set. The pre-training of the adaptive fuzzy prediction model based on collaborative decision-making is completed.

[0149] Furthermore, determining all the fuzzy rules and the corresponding conclusion parameter groups of the adaptive fuzzy prediction model based on collaborative decision-making according to the historical transformation data set includes the following specific steps:

[0150] Obtain the training set and the test set in the historical transformation data set, generate 1 fuzzy rule for the first input historical transformation sample in the training set, and use the variance of the historical transformation samples in the training set as the initial width of the first Gaussian membership function Take the first input historical transformation sample as the initial center of the first Gaussian membership function Generate the first fuzzy set;

[0151] Input the second historical transformation sample in the training set, fuzzify it according to the existing Gaussian membership function, and calculate the strength g corresponding to the second historical transformation sample under the existing fuzzy rules 1 , And compare it with the set coverage threshold;

[0152] If the strength g corresponding to the second historical transformation sample under the existing fuzzy rules 1If it is less than the set coverage threshold, the second fuzzy rule is generated, and the variance after removing the first historical transformation sample from the training set is used as the initial width of the second Gaussian membership function. The historical transformation sample of the second input is used as the initial center of the second Gaussian membership function. Generate the second fuzzy set and continue to input the third historical transformation sample in the training set;

[0153] If the intensity g corresponding to the second historical transformation sample under the existing fuzzy rules 1 is greater than or equal to the set coverage threshold, no new fuzzy rules, new Gaussian membership functions, and new fuzzy sets are added, and the third historical transformation sample in the training set is continued to be input;

[0154] Repeat the above logic until the entire training set is traversed, and a total of K fuzzy rules, K Gaussian membership functions, and K fuzzy sets are obtained. Among them, the fuzzy rules and Gaussian membership functions correspond one-to-one with the fuzzy sets;

[0155] According to the linear combination form of the total input factors f 1 ,…,f K adopt the least squares method to solve the initial conclusion parameter group corresponding to the K fuzzy rules

[0156] Input the test set into the collaborative decision-making based adaptive fuzzy prediction model for rule verification and calculate the F1-Score of the test set. Further, use the adaptive optimization algorithm to adjust the widths and centers of the K Gaussian membership functions and the conclusion parameter groups corresponding to the K fuzzy rules to improve the F1-Score of the test set. Finally, determine the widths and centers of the K Gaussian membership functions corresponding to the historical transformation data set and the conclusion parameter groups corresponding to the K fuzzy rules. Among them, the adaptive optimization algorithm includes the grey wolf optimization algorithm, whale foraging algorithm, quantum genetic algorithm, particle swarm algorithm, and backpropagation algorithm.

[0157] Based on the following specific case, illustrate the inference logic of the collaborative decision-making based adaptive fuzzy prediction model. In the time period from 0:00 to 5:00, the ventilation state of the working face in the coal mine is basically stable, and the wind speed x w always remains near 0.87 m / s, that is, the wind speed change rate is approximately 0. At 0:00, the gas concentration x in the roadway c is 0.34%, and the stress intensity gx s is equal to 0, and the working face is in a stress balance state;

[0158] In the time period from 0:15 to 02:15, the stress intensity gx of the working face sGreater than 0 but less than 1, the first concentrated stress impact phenomenon occurs, lasting for 22 minutes, and the impact energy peak value x e (t) reaches 4976×104 joules, and the gas concentration after impact is x c A decline occurred, down to 0.26%, but the impact energy was not enough to destroy the coal seam structure, only the coal seam gas was compressed and gathered. During this process, the adaptive fuzzy prediction model based on collaborative decision-making determined that the outburst warning level y = 1;

[0159] During the period of 2:18 to 2:40, the second concentrated stress impact phenomenon occurred on the working face, which lasted for 15 minutes and the peak impact energy was x e (t) reaches 19417×104 joules / min, which is more than twice the first impact force. The gas concentration x c It rose to 0.31%, which means that the second stress impact crushed the coal seam within a certain range, increasing the permeability of the coal seam, and the adsorbed gas began to transform into free gas, which gathered in the coal seam to form gas expansion energy. In this process, the adaptive fuzzy prediction model based on collaborative decision-making determined that the warning level y changed from 1 to 2;

[0160] During the period from 2:40 to 5:00, the gas concentration continued to rise for about 25 minutes. At 3:02, the gas concentration was x c Reach 0.78%, but the stress intensity gx of the working surface s It is always less than 1, the impact force formed by the concentration is small, the range of the crushed coal seam is small, the gas expansion energy formed is insufficient and weak, and it has not developed into a real outburst. After 3:06, the gas concentration x c It begins to decline gradually. During this process, the prominent warning level y determined by the adaptive fuzzy prediction model based on collaborative decision-making changes from 2 to 1 and then to 0.

[0161] Specifically, the knowledge graph needs to be pre-built and continuously updated. The construction of the knowledge graph includes the following specific steps:

[0162] The application scenarios of the knowledge graph are clarified as mine safety monitoring, accident prevention and emergency response, and the model layer of the knowledge graph is defined, including entities, relationships and attributes, such as mine, equipment, accident type and safety specification entities;

[0163] Collect mine safety data in advance, and clean, denoise and format the data. Mine safety data includes structured data, semi-structured data and unstructured data. Structured data is database records, semi-structured data is web pages and XML files, and unstructured data is text and images.

[0164] Use NLP techniques for text tokenization, part-of-speech tagging, named entity recognition, and relation extraction to identify mine safety entities, attributes, and relations in the text. Among them, entities include equipment names, accident types, personnel distribution, emergency plans, and historical cases; attributes include equipment models and accident levels; relations include equipment-failure and accident-cause.

[0165] Apply image recognition methods to identify mine safety elements in images, such as equipment status and safety hazards, and extract text information in images, such as safety signs and equipment labels.

[0166] Fuse the knowledge extracted from text and images to solve entity alignment and relation conflict problems, and store the fused knowledge in a graph database in the form of Neo4j for efficient querying and reasoning.

[0167] Combine mine GIS technology to obtain the position of the roadheader. Based on the knowledge graph and rule engine, comprehensively analyze the personnel status z(t) and outburst warning level y(t) at time t, automatically match the optimal emergency plan and execute it. Store the clear image J(t), environmental information X(t), outburst warning level y(t), roadheader position, and optimal emergency plan in the knowledge graph to update the knowledge graph.

[0168] Specifically, the rule engine analyzes the personnel status z(t) and outburst warning level y(t) based on conditional judgment logic, and calls the knowledge of the knowledge graph to automatically formulate the optimal emergency plan. The conditional judgment logic includes:

[0169] If the personnel status z(t) is actionable and the outburst warning level y(t)=0, it indicates that the staff is in a normal state and there is no outburst risk, and the matched optimal emergency plan is empty.

[0170] If the personnel status z(t) is actionable and the outburst warning level y(t)=1, it indicates that the staff is in a normal state and there is a general outburst risk. The matched optimal emergency plan is to continue intermittent operation and send it to the staff, and control the audible and visual alarm on the roadheader to display blue light wirelessly. Among them, intermittent operation specifically means reducing the roadheader's tunneling speed, thereby reducing the degree of concentration stress aggregation and reducing the risk of outburst occurrence.

[0171] If the personnel status z(t) is actionable and the outburst warning level y(t)=2, it indicates that the staff is in a normal state and there is a serious outburst risk. The matched optimal emergency plan is to stop the operation, combine mine GIS technology and heuristic algorithm to determine the optimal evacuation route from the personnel position to the mine exit, feedback the optimal emergency plan and optimal evacuation route to the staff, and control the audible and visual alarm on the roadheader to display orange light wirelessly.

[0172] If the personnel status z(t) is immobile, it indicates that the status of the staff is abnormal. The optimal emergency plan is to stop the operation and dispatch rescue. A stop signal is sent through wireless transmission to remotely control the roadheader to stop working. The optimal rescue route from the mine exit to the personnel location is determined by combining the mine GIS technology and the heuristic algorithm and sent to the rescue personnel.

[0173] As Figure 3 shown, further, determining the optimal evacuation route or the optimal rescue route by combining the mine GIS technology and the heuristic algorithm includes the following specific steps:

[0174] Restore the three-dimensional mine map based on the mine GIS technology. The three-dimensional mine map shows each mine tunnel, marks the position of each roadheader at time t and the corresponding outburst warning level y(t);

[0175] Use a three-dimensional grid to divide the three-dimensional mine map into multiple regional grids, remove the regional grids corresponding to the outburst warning level y(t)=2, and attach a danger mark to the adjacent neighbor regional grids. Determine the starting regional grid and the target regional grid. Among them, the starting regional grid and the target regional grid of the optimal evacuation route and the optimal rescue route are exactly opposite;

[0176] Select the A-Star algorithm for path planning. The A-Star algorithm establishes a parent list and a child list. Place the starting regional grid into the parent list. The parent list stores the explored regional grids, and the child list stores the unexplored regional grids;

[0177] Starting from the starting regional grid, at each step of the A-Star algorithm, calculate the evaluation value of each neighbor regional grid adjacent to the current regional grid in the child list, and select the neighbor regional grid with the highest evaluation value and place it into the parent list. Among them, the evaluation value is equal to the Manhattan distance from the neighbor regional grid to the target regional grid minus the penalty term. The penalty term is the first fixed value or the second fixed value. If the neighbor regional grid has a danger mark, the penalty term is the first fixed value. If the neighbor regional grid has no danger mark, the penalty term is the second fixed value. In this embodiment, the first fixed value and the second fixed value are selected as 20 and 0;

[0178] Judge whether the neighbor regional grid is the target regional grid. If it is not the target regional grid, use the neighbor regional grid as the current regional grid and continue to calculate the evaluation values of all neighbor regional grids of the current regional grid to determine the next neighbor regional grid to go to;

[0179] If it is the target regional grid, generate the optimal evacuation route or the optimal rescue route according to the storage order of the regional grids in the parent list.

[0180] Embodiment 2

[0181] As Figure 4As shown in the figure, the present invention also discloses a prominent danger early warning analysis system suitable for application on a roadheader, which is used to implement a prominent danger early warning analysis method suitable for application on a roadheader and can communicate wirelessly with a prominent warning electric control platform installed on the roadheader. The system includes a wireless communication module, a comprehensive early warning module, an emergency response module, and a storage module;

[0182] The wireless communication module receives the image I and environmental information X periodically collected by the acquisition device of the prominent warning electric control platform installed on the roadheader body through wireless communication means. The acquisition period is equal to the interval duration Δt. The acquisition device includes a front camera for taking the coal seam image I of the working face s of the working face, a rear camera for taking the personnel image I p a gas sensor for measuring the gas concentration x c a stress sensor for measuring the coal seam stress x s a wind speed sensor for measuring the wind speed x w and four microseismic sensors for monitoring microseismic signals;

[0183] The comprehensive early warning module preprocesses the image I by using the multiple transmission enhancement method to generate a clear image J according to the image I and environmental information X provided by the wireless communication module, and obtains the personnel state z, soft rock bearing strength and hard rock bearing strength by using the target state recognition network and the coal seam rock mass recognition network, calculates the gas concentration change rate the stress intensity gx s and the wind speed change rate and performs fuzzy reasoning by using an adaptive fuzzy prediction model based on collaborative decision-making to generate a prominent danger early warning level y;

[0184] The emergency response module combines the mine GIS technology to obtain the position of the roadheader, comprehensively analyzes the personnel state z and the prominent danger early warning level y based on the knowledge graph and the rule engine, matches and feeds back the optimal emergency plan to the staff or rescue personnel, and at the same time issues an audible and visual alarm prompt;

[0185] The storage module stores the knowledge graph and dynamically updates the knowledge graph according to the clear image J, environmental information X, prominent danger early warning level y, roadheader position, and optimal emergency plan.

[0186] Furthermore, the comprehensive early warning module includes a data processing unit, an image analysis unit, and a risk assessment unit;

[0187] The data processing unit weakens the light source and performs multi-transmission fusion on the image I by using the multiple transmission enhancement method, and generates a clear image J based on the atmospheric scattering model to solve the problem of image I blurring caused by the dim mine environment;

[0188] The image analysis unit captures a clear image of personnel J through the target state recognition network p generates the personnel state z based on the postures of the staff in p , and precisely analyzes the rock color, texture, and the contact relationship between the coal seam and the roof and floor in the clear working face coal seam image at time t through the coal seam rock mass recognition network, identifies the type of supporting rock mass, and determines the bearing strength of soft rock and the bearing strength of hard rock The supporting rock mass includes soft rock and hard rock;

[0189] The risk assessment unit calculates the change rate of gas concentration the stress intensity gx s and the change rate of wind speed Adopts an adaptive fuzzy prediction model based on collaborative decision-making, considering domain experience and mathematical analysis, the change rate of gas concentration the stress intensity gx s the change rate of wind speed the peak value of impact energy x e and the inherent rate x j are transformed into an input factor sequence Based on the change correlation between input factors, fuzzy reasoning is carried out, and defuzzification is performed to generate the outburst warning level y.

[0190] As an example, a method and system for outburst danger warning analysis provided by the present invention, which are suitable for application on a roadheader, can be applied in a outburst prevention roadheader with the function of warning coal and gas outburst danger as shown in Figure 5 The outburst prevention roadheader with the function of warning coal and gas outburst danger includes a roadheader assembly 100, a roadheader body 101, and a warning assembly 200 provided at the end face of the roadheader body 101; the warning assembly 200 is installed at a position on the rear side of the roadheader body 101 to reduce the vibration influence of the front drill bit during the cutting and tunneling operations of the roadheader, and at the same time reduce the risk of being hit by the crushed stones generated during the tunneling of the front drill bit. A outburst warning electronic control platform is provided at the end of the roadheader body 101, and the outburst warning electronic control platform includes a main circuit board assembly and a data acquisition assembly; the main circuit board assembly is composed of seven single-chip microcomputers, which are used to control the periodic acquisition of the data acquisition assembly and provide wireless communication functions; the data acquisition assembly acquires environmental information to provide data support for the coal and gas outburst warning, and at the same time obtains the working face coal seam image and the personnel image through the front camera and the rear camera respectively.

[0191] The present invention discloses a method and system for outburst danger early warning analysis suitable for application on a roadheader. It receives real-time images and environmental information collected by an outburst early warning electronic control platform through a wireless transmission method, and preprocesses the images through a multiple transmission enhancement method to solve the problem of image blurring caused by a dim environment, assisting the target state recognition network and the coal seam rock mass recognition network to obtain accurate personnel states and determine the bearing strength of soft rock and hard rock, calculating the change rate of gas concentration, stress intensity, and wind speed change rate, and converting them together with the peak impact energy and the inherent rate into input factors and inputting them into an adaptive fuzzy prediction model based on collaborative decision-making. Reasoning is carried out according to fuzzy rules considering the change correlation of the input factors, and defuzzification is performed to generate an outburst early warning level. Combining the mine GIS technology and comprehensively analyzing the personnel state and the outburst early warning level based on a knowledge graph and a rule engine, automatically matching and executing the optimal emergency plan and updating the knowledge graph, realizing the real-time monitoring of outburst danger in the mechanical tunneling face and the rapid execution of the emergency plan, and improving the safety of mine operations and the emergency response efficiency.

[0192] The above is only the preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the idea of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art in this technical field, several improvements and refinements made without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A method for early warning analysis of outburst danger suitable for application on a tunnel boring machine, characterized in that: The specific steps include: Obtain the coal seam images, personnel images, gas concentration, coal seam stress, wind speed, impact energy peak and co-occurrence rate of the working face collected by the acquisition equipment of the outburst warning electric control platform through wireless communication; Preprocess the personnel image and identify the personnel status by capturing the personnel posture features through the target state recognition network; preprocess the coal seam image of the working face and analyze the rock texture and color through the coal seam rock mass recognition network, identify the soft rock and hard rock types and determine the bearing strength of the soft rock and hard rock; Calculate the gas concentration change rate and wind speed change rate, compare the coal seam stress with the bearing strength of soft rock and hard rock to calculate the stress intensity, and multiply the gas concentration change rate, stress intensity, wind speed change rate, impact energy peak and co-occurrence rate by the corresponding input factor weights based on the adaptive fuzzy prediction model of collaborative decision-making to construct an input factor sequence. According to the fuzzy rules, the change correlation of the input factors is considered for reasoning and defuzzification to generate the outburst warning level. Combined with mine GIS technology, the location of the tunnel boring machine is obtained, and the optimal emergency plan is executed based on the knowledge graph and rule engine to match the personnel status and prominent warning level, and the knowledge graph is updated.

2. A method for early warning analysis of outburst danger suitable for application on a tunnel boring machine as claimed in claim 1, characterized in that: The generation of the prominent warning level comprises the following specific steps: The gas concentration change rate, stress intensity, wind speed change rate, and impact energy peak value and coincidence rate in the environmental information are multiplied by the corresponding input factor weights to generate an input factor sequence; The membership of a single input factor and each fuzzy set is calculated using the Gaussian membership function to generate a membership vector corresponding to the single input factor. The product of each input factor and the membership degree of a single fuzzy set is taken as the strength of the corresponding fuzzy rule, the strength of each fuzzy rule is obtained and normalized to generate the corresponding strength weight; According to the conclusion parameter group of a single fuzzy rule, all input factors are linearly summed to generate the corresponding total input factor. According to the intensity weights of all fuzzy rules, the corresponding total input factors are weighted and summed to generate the prominent warning level.

3. A method for early warning analysis of outburst danger suitable for application on a roadheader as claimed in claim 1 or 2, characterized in that: The adaptive fuzzy prediction model based on collaborative decision-making requires pre-training, which includes the following specific steps: Calculate the probability of being selected for the historical gas concentration change rate, stress intensity, wind speed change rate, impact energy peak value and co-occurrence rate; Obtain historical data sets, and calculate the Gini values ​​of the gas concentration change rate, stress intensity, wind speed change rate, impact energy peak value, and common rate separately classified historical data sets; Initialize multiple sets of first ratios and second ratios and process the historical data sets respectively to generate multiple different historical transformation data sets, and divide each historical transformation data set into a training set and a test set; For the training set of a single historical transformation data set, historical transformation samples are input in sequence, the strength of the existing fuzzy rules is calculated based on the existing Gaussian membership function, and it is compared with the coverage threshold to decide whether to generate new fuzzy rules and Gaussian membership functions. After traversing the training set, the least squares method is used to solve the initial conclusion parameter group of each fuzzy rule; The evaluation index is calculated through the test set of a single historical transformation data set, and the parameters of the Gaussian membership function and the initial conclusion parameter group generated by the corresponding training set are adjusted by an adaptive optimization algorithm to optimize the evaluation index. The parameters of the Gaussian membership function and the initial conclusion parameter group are fixed when the algorithm converges. All fuzzy rules, the Gaussian membership function of each fuzzy rule and the conclusion parameter group are fixed by selecting the historical transformation data set with the best evaluation index.

4. A method for early warning analysis of outburst danger suitable for application on a tunnel boring machine as claimed in claim 1, characterized in that: The calculation of the gas concentration change rate, stress intensity, and wind speed change rate includes: Obtain environmental information such as gas concentration, coal seam stress and wind speed; The gas concentration change rate is the change obtained by subtracting the gas concentration at the previous moment from the current moment’s gas concentration divided by the interval duration; The stress intensity is equal to 0 when the coal seam stress at the current moment is less than or equal to the soft rock bearing strength. When the coal seam stress at the current moment is greater than the soft rock bearing strength, the stress intensity is equal to the difference between the coal seam stress at the current moment and the soft rock bearing strength divided by the difference between the hard rock bearing strength and the soft rock bearing strength. The wind speed change rate is the change obtained by subtracting the wind speed at the previous moment from the current moment divided by the interval length.

5. A method for early warning analysis of outburst danger suitable for application on a tunnel boring machine as claimed in claim 1, characterized in that: The knowledge graph needs to be constructed in advance, and the construction of the knowledge graph includes the following specific steps: Construct application scenarios of knowledge graphs and define the model layer of knowledge graphs; Collect mine safety data in advance and perform cleaning, denoising and formatting; Use NLP technology to perform text segmentation, part-of-speech tagging, named entity recognition and relationship extraction to identify mine safety entities, attributes and relationships in the text; Apply image recognition methods to identify mine safety elements in images and extract text information from images; The knowledge extracted from text and image knowledge is fused and stored in the graph database in the form of Neo4j.

6. A method for early warning analysis of outburst danger suitable for application on a tunnel boring machine as claimed in claim 1, characterized in that: The method of executing the optimal emergency plan for matching personnel status and prominent warning level based on knowledge graph and rule engine includes the following specific steps: If the personnel status is ready for action and the prominent warning level is 0, the matching optimal emergency plan is empty; If the personnel status is ready for action and the prominent warning level is 1, the optimal emergency plan is intermittent operation and sent to the staff; If the personnel status is ready for action and the outburst warning level is 2, the optimal emergency plan is to stop working. Combine the mine GIS technology and heuristic algorithm to determine the optimal evacuation route from the location to the mine exit and feedback to the staff; If the personnel status is incapable of action, the optimal emergency plan is to stop the operation and dispatch rescue. The optimal rescue route from the mine exit to the location is determined by combining mine GIS technology and heuristic algorithms and sent to the rescue personnel.

7. A method for early warning analysis of outburst danger suitable for application on a tunnel boring machine as claimed in claim 1, characterized in that: The preprocessing of the personnel image and the working face coal seam image adopts a multiple transmission enhancement method, including the following specific steps: The image is converted into an HSV image, and the brightness map on the brightness channel V is minimum filtered to change the brightness value of any point in the image to the minimum brightness of all points in the local neighborhood centered on the point, and a weakened brightness map is obtained. The image includes a personnel image and a coal seam image of the working face; Take the maximum brightness value in the weakened brightness image as the atmospheric light value; According to the dark channel principle, the darkness value of any point in the image at each order is converted into the minimum color component of all points in the local block centered at the point and with a scale equal to the order in the RGB three-color channels, and dark images of all orders are obtained; The transmittance of each order of dark image is calculated based on the linear relationship between the ratio of the dark image to the atmospheric light value and the transmittance, and the average is taken as the transmittance of the image; Combined with the atmospheric scattering model, clear images are generated through atmospheric light values ​​and transmittance. The clear images include clear personnel images and clear working face coal seam images.

8. A method for early warning analysis of outburst danger suitable for application on a roadheader as claimed in claim 1, characterized in that: The target state recognition network is used to obtain the state of the personnel, including the following specific steps: Extract the preprocessed person image through continuous normalized convolution to generate different levels of image information; The residual structure is used to superimpose and fuse image information of different levels to generate multi-scale image information; The spatial features of people in multi-scale image information are filtered through the spatial attention mechanism and the correlation between the spatial features of people and spatial posture is strengthened to obtain the spatial feature map of people. Through the channel attention mechanism, we focus on the pixel points of multi-scale image information from the three RGB channels to generate a semantic feature map of the person; The personnel spatial feature map and the personnel semantic feature map are superimposed as the real part and the imaginary part respectively, and classified by the complex linear layer and the complex support vector machine in turn to generate the personnel status, which includes actionable and inaction.

9. A method for early warning analysis of outburst danger suitable for application on a roadheader as claimed in claim 1, characterized in that: The determination of the bearing strength of soft rock and hard rock comprises the following specific steps: The preprocessed coal seam image of the working face is subjected to deep convolution on the three RGB channels respectively, and the working face feature map is generated by point-by-point convolution combination; The working surface feature map is subjected to multiple dilated convolutions with different dilation rates to generate multiple extended feature maps, and then point-by-point convolutions are applied again to generate an aggregated extended feature map. The aggregated extended feature map is sequentially subjected to bilinear interpolation, 4-fold upsampling and feature compression to achieve semantic feature fusion and generate a coal-rock fusion feature map. The coal-rock fusion feature map is input into two Softmax logistic regressors for linear dimension reduction and nonlinear activation to generate soft rock type probability vector and hard rock type probability vector; The soft rock bearing strength and the hard rock bearing strength are determined according to the soft rock type and the hard rock type with the maximum probability in the soft rock type probability vector and the hard rock type probability vector.

10. A hazard early warning analysis system suitable for use on a tunnel boring machine, characterized in that: It includes wireless communication module, comprehensive warning module, emergency response module and storage module; The wireless communication module periodically receives images and environmental information collected and transmitted by the outburst warning electronic control platform; The comprehensive warning module uses a multiple transmission enhancement method to pre-process images to generate clear images, obtains personnel status, soft rock bearing strength and hard rock bearing strength through a target state recognition network and a coal seam rock mass recognition network, calculates the gas concentration change rate, stress intensity and wind speed change rate, and uses an adaptive fuzzy prediction model based on collaborative decision-making to perform fuzzy reasoning to generate a sudden warning level; The emergency response module combines the mine GIS technology to obtain the location of the tunnel boring machine, comprehensively analyzes the personnel status and the level of outstanding warning based on the knowledge graph and rule engine, and matches the optimal emergency plan; The storage module stores the knowledge graph and updates it dynamically.

Images