Ventilation data analysis method and system for intelligent ventilation of mine

Abstract

The invention belongs to the technical field of mine ventilation data analysis, and particularly provides a ventilation data analysis method and system for mine intelligent ventilation, and the method mainly comprises the steps: obtaining mine environment parameters which comprise gas concentration, real-time air volume and temperature; calculating a dynamic risk index according to the mine hole environment parameters; generating a risk trend state code according to the dynamic risk index; and an atmospheric pressure sequence in a past specified time range is obtained, the atmospheric pressure sequence is processed by adopting a self-adaptive atmospheric pressure model, and the self-adaptive atmospheric pressure model takes the time convolutional network as a basic framework to dynamically determine the optimal expansion rate of the expansion convolutional layer. According to the method, the environmental evolution law is quantized based on real-time risk trend coding, the multi-scale air pressure characteristics extracted in a self-adaptive mode are combined, meanwhile, the dynamic association of the tunneling state and the risk trend is integrated, the coupling influence of mechanical disturbance and air pressure fluctuation can be effectively decoupled, gas concentration abnormal fluctuation is early warned in advance, and the safety of the gas concentration is improved. And the safety prevention and control real-time performance under the complex working condition is obviously improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of mine ventilation data analysis, and in particular relates to a ventilation data analysis method and system for mine intelligent ventilation. Background Art

[0002] When mining, harmful gases such as gas will be produced, and mechanical operations will bring dust. The narrow space can easily cause the air to be stuffy and oxygen-deficient. Mine ventilation can expel harmful gases and dust, replenish fresh air, lower the temperature, create a safe and comfortable working environment for miners, prevent dangers such as gas accumulation and explosion, and ensure smooth production.

[0003] The dynamic imbalance of mine ventilation systems stems from the multi-source coupling of gas, air pressure and mechanical disturbances. Nonlinear time-varying correlations cause ventilation forecasts to continuously deviate from actual demand, inducing local gas over-limit and energy consumption surges. Traditional methods rely on static models and fixed time windows to analyze air pressure data, which conceals the interaction between short-term mutations and long-term trends, and makes it difficult to capture the dynamic correlation between multi-scale fluctuations in air pressure and risk evolution, making it difficult to analyze the superposition of mechanical disturbances on time series characteristics. It is difficult to correct parameter coupling deviations under complex working conditions, which easily leads to safety prevention and control lagging behind risk evolution. Summary of the invention

[0004] The present invention provides a ventilation data analysis method and system for intelligent mine ventilation, which effectively solves the problem that the prior art relies on static models and fixed time windows to analyze air pressure data, which easily leads to safety control lagging behind the evolution of risks. It can provide early warning of abnormal fluctuations in gas concentration and significantly improve the real-time safety control under complex working conditions.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions: In a first aspect, the present disclosure provides a ventilation data analysis method for intelligent ventilation in mines, comprising: obtaining mine environment parameters, which include gas concentration, real-time air volume and temperature; calculating a dynamic risk index based on the mine environment parameters; generating a risk trend state code based on the dynamic risk index; obtaining an atmospheric pressure sequence within a specified time range in the past, and processing the atmospheric pressure sequence using an adaptive atmospheric pressure model, wherein the adaptive atmospheric pressure model is based on a time convolutional network, dynamically determines the optimal expansion rate of an expansion convolution layer, and outputs a time series feature vector after deep integration of multi-scale information; obtaining the excavation status of a tunnel boring machine; and obtaining a ventilation volume prediction interval for a specified time range in the future based on the time series feature vector, the risk trend state code, the excavation status of the tunnel boring machine and the real-time air volume.

[0006] Furthermore, the mine environment parameters are obtained, including: configuring basic acquisition parameters, including: setting the sampling frequency of gas concentration, real-time air volume and temperature, defining the time window and the space window, and the sliding interval of the time window and the sliding interval of the space window; based on the basic acquisition parameters, according to the sliding of the time window, a three-dimensional time series tensor including a time axis, spatial node distribution and parameter channels is generated; according to the sliding of the spatial window, the three-dimensional time series tensor is dynamically trimmed to obtain valid node data.

[0007] Furthermore, defining the spatial window includes: taking the tunnel boring machine as a reference point, using an ultra-wideband positioning system to obtain the real-time coordinates of the reference point; determining the coordinates of the sensor, calculating the actual distance between the sensor and the reference point; and screening out sensors whose actual distances meet the preset distances.

[0008] Furthermore, a dynamic risk index is calculated based on the mine environmental parameters, including: for valid node data in the current time window and space window: calculating the mean gas concentration based on the sampling frequency and spatial node distribution of the gas concentration; normalizing the real-time air volume, and calculating the mean of the normalized real-time air volume based on the sampling frequency and spatial node distribution of the real-time air volume; calculating the mean temperature based on the sampling frequency and spatial node distribution of the temperature; calculating the gas concentration risk factor, air volume risk factor and temperature risk factor based on the mean gas concentration, the mean normalized real-time air volume and the mean temperature; calculating the dynamic risk index based on the gas concentration risk factor, the air volume risk factor and the temperature risk factor.

[0009] Furthermore, a risk trend status code is generated based on the dynamic risk index, including: determining two or more continuously increasing risk levels, each risk level corresponding to a unique code; defining a demarcation threshold value that is one less than the number of risk levels, and all demarcation threshold values ​​are arranged in ascending order; determining the demarcation threshold range in which the dynamic risk index is located, and outputting the risk level code according to the code mapping mechanism.

[0010] Among them, the coding mapping mechanism includes: mapping the dynamic risk index below the minimum demarcation threshold to the lowest risk level code; mapping the dynamic risk index between every two adjacent demarcation thresholds to the intermediate risk level code in the order of the demarcation thresholds; mapping the dynamic risk index higher than or equal to the maximum demarcation threshold to the highest risk level code.

[0011] Furthermore, the adaptive atmospheric pressure model includes an input layer, a first specified number of dilated convolutional layers and an output layer; the input layer is used to receive the atmospheric pressure sequence within a specified time range in the past; the first specified number of dilated convolutional layers are assigned an optimal dilation rate to capture the changing characteristics of atmospheric pressure under different periods, obtain multiple feature maps, splice the feature maps by channels, and integrate multi-scale information into a feature tensor; the output layer flattens the feature tensor to obtain a time series feature vector.

[0012] Further, in the process of assigning the optimal expansion rate to the first specified number of expansion convolutional layers, obtaining the optimal expansion rate includes: obtaining an atmospheric pressure sequence within a specified time range in the past; performing multi-scale decomposition on the atmospheric pressure sequence to obtain multiple subsequences; extracting statistical features of the subsequences to obtain feature vectors; clustering the feature vectors of all subsequences to obtain a second specified number of cluster centers and a cluster label to which each subsequence belongs; classifying an initial expansion rate for each cluster center to obtain an initial expansion rate set; when training an adaptive atmospheric pressure model, calculating the importance weight of the subsequence corresponding to each cluster in the current input; fine-tuning the initial expansion rate according to the importance weight of the subsequence corresponding to each cluster in the current input to obtain a final expansion rate; constructing a set of final expansion rates, using the set of final expansion rates to train the adaptive atmospheric pressure model and perform feedback adjustment to obtain the optimal expansion rate.

[0013] Furthermore, based on the time series feature vector, risk trend state code, tunneling state of the tunnel boring machine and real-time air volume, a ventilation volume prediction interval for a specified time range in the future is obtained, including: using the GBRT improved architecture to process the time series feature vector, risk trend state code, tunneling state of the tunnel boring machine and real-time air volume, and outputting the ventilation volume prediction interval for a specified time range in the future.

[0014] Among them, the GBRT improved architecture includes: input layer, feature fusion layer, gradient boosting tree set layer and output layer; the input layer is used to receive time series feature vectors, risk trend state codes, tunneling status of the tunnel boring machine and real-time air volume; the feature fusion layer is used to splice the time series feature vectors and the risk trend state code vectors by channel to form fusion features; the real-time air volume is subjected to sliding window mean filtering to generate air volume trend features; the fusion features, air volume trend features and tunneling status parameters are spliced ​​into a joint feature vector; the gradient boosting tree set layer contains a third specified number of regression trees, each regression tree performs node splitting based on the joint feature vector, and outputs the ventilation volume prediction base value for each time step; the output layer is used to perform quantile regression on the ventilation volume prediction base value, and the quantile loss function is used to calculate the upper quantile and lower quantile of the future specified time range to generate the ventilation volume prediction interval.

[0015] Furthermore, the quantile output layer dynamically scales the upper and lower bounds of the ventilation volume prediction interval according to the ratio of the current dynamic risk index to the historical risk interval, and the scaling coefficient is positively correlated with the dynamic risk index.

[0016] In a second aspect, the present disclosure provides a ventilation data analysis system for intelligent ventilation in mines, which includes: a multi-source environmental parameter perception module, a dynamic risk assessment and fusion module, a risk coding and air pressure time series feature generation module and a multimodal ventilation decision module.

[0017] The multi-source environmental parameter perception module is used to obtain the mine environment parameters, which include gas concentration, real-time air volume and temperature; the dynamic risk assessment and fusion module is used to calculate the dynamic risk index based on the mine environment parameters; the risk coding and air pressure time series feature generation module is used to generate the risk trend state code based on the dynamic risk index; the atmospheric pressure sequence within the specified time range in the past is obtained, and the atmospheric pressure sequence is processed by an adaptive atmospheric pressure model. The adaptive atmospheric pressure model is based on the time convolution network, dynamically determines the optimal expansion rate of the expansion convolution layer, and outputs the time series feature vector after deep integration of multi-scale information; the multimodal ventilation decision module is used to obtain the excavation status of the tunnel boring machine; according to the time series feature vector, risk trend state code, tunnel boring machine excavation status and real-time air volume, the ventilation volume prediction interval for the specified time range in the future is obtained.

[0018] In a third aspect, the present disclosure provides a device comprising a memory and a processor; the memory is used to store a computer program; the processor is used to implement the steps of the ventilation data analysis method for intelligent ventilation in mines as described in the first aspect when executing the computer program.

[0019] In a fourth aspect, the present disclosure provides a storage medium storing computer program instructions. When the computer program instructions are read and executed by a processor, the steps of the ventilation data analysis method for intelligent ventilation in mines as described in the first aspect are executed.

[0020] Beneficial effects of the present invention: The present invention quantifies the environmental evolution law based on real-time risk trend coding, combines the multi-scale air pressure characteristics extracted adaptively, and integrates the dynamic relationship between excavation status and risk trend. It effectively solves the problem that the existing technology relies on static models and fixed time windows to analyze air pressure data, which easily leads to safety prevention and control lagging behind the risk evolution. It can effectively decouple the coupling effects of mechanical disturbances and air pressure fluctuations, give early warning of abnormal fluctuations in gas concentration, and significantly improve the real-time safety prevention and control under complex working conditions.

[0021] Other features and advantages of the present invention will be described in the following description, and partly become obvious from the description, or be understood by implementing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0023] Figure 1 A schematic flow chart of a ventilation data analysis method for intelligent ventilation of a mine according to the present invention is shown; Figure 2 A module schematic diagram of a ventilation data analysis system for intelligent mine ventilation of the present invention is shown. DETAILED DESCRIPTION

[0024] In order to solve the problems raised by the background technology, the present invention quantifies the environmental evolution law based on real-time risk trend coding, combines the multi-scale gas pressure characteristics extracted adaptively, and integrates the dynamic correlation between excavation status and risk trend to provide early warning of abnormal fluctuations in gas concentration, thereby significantly improving the real-time safety prevention and control under complex working conditions.

[0025] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] In some embodiments, Figure 1 As shown, the present disclosure provides a ventilation data analysis method for intelligent ventilation in a mine, comprising: S100. Obtaining mine environment parameters, which include gas concentration, real-time air volume and temperature.

[0027] S200. Calculate the dynamic risk index based on the mine environment parameters.

[0028] S300. Generate a risk trend state code according to the dynamic risk index; obtain the atmospheric pressure sequence within a specified time range in the past, and use an adaptive atmospheric pressure model to process the atmospheric pressure sequence. The adaptive atmospheric pressure model is based on a time convolutional network, dynamically determines the optimal expansion rate of the expansion convolution layer, and outputs a time series feature vector after deep integration of multi-scale information.

[0029] S400. Obtain the excavation status of the tunnel boring machine; obtain the ventilation volume prediction interval within a specified time range in the future according to the time series feature vector, the risk trend state code, the excavation status of the tunnel boring machine and the real-time air volume.

[0030] In some embodiments, obtaining the mine environment parameters in S100 includes: S110. Configure basic collection parameters, including: setting the sampling frequency of gas concentration, real-time air volume and temperature, defining the time window and space window, and the sliding interval of the time window and the sliding interval of the space window.

[0031] Set the sampling frequency of gas concentration, real-time air volume and temperature. For example, the sampling frequency of gas concentration is once per second, the sampling frequency of real-time air volume is twice per second, and the sampling frequency of temperature is once per second.

[0032] Define the size and sliding interval of the time window and space window: For example, set the time window length to 10 minutes and the time window sliding interval to 5 minutes, that is, the window slides forward every 5 minutes, and the window contains data for 10 consecutive minutes.

[0033] For example, with the current position of the tunnel boring machine as the center, the spatial window is defined as a circular area with a radius of 5 meters, and the sliding interval of the spatial window is 1 meter, that is, the spatial window is repositioned every time the tunnel boring machine moves 1 meter.

[0034] S120. Based on the basic acquisition parameters, a three-dimensional time series tensor including a time axis, spatial node distribution, and parameter channels is generated according to the sliding of the time window.

[0035] Timeline construction: Generate a time step sequence at the sampling frequency within each time window. For example, a 10-minute time window contains 600 time steps.

[0036] Spatial node distribution: In the spatial window, the coordinates of each sensor are obtained through the ultra-wideband positioning system to generate a spatial node distribution matrix. For example, if there are 5 sensors in the current spatial window, the number of spatial nodes is 5.

[0037] Each spatial node collects three parameters, namely, gas concentration, real-time air volume, and temperature, at each time step to form a parameter channel dimension.

[0038] The final generated three-dimensional time series tensor structure is: tensor shape = (time step, number of spatial nodes, 3).

[0039] S130. Dynamically crop the three-dimensional time series tensor according to the sliding of the spatial window to obtain valid node data.

[0040] S131. Obtain the real-time coordinates of the tunnel boring machine according to the ultra-wideband positioning system, and calculate the Euclidean distance between each sensor and the tunnel boring machine : ;in,( , , ) represents the coordinates of the tunnel boring machine, ( , , ) represents the coordinates of the i-th sensor.

[0041] S132. Filter out the Euclidean distance For sensor nodes ≤5 meters away, a valid node index list is generated.

[0042] S133. According to the valid node index list, extract the data of the corresponding spatial nodes from the original three-dimensional time series tensor to obtain the valid node data.

[0043] For example, the shape of the original three-dimensional time series tensor is 600×20×3. After filtering by the spatial window, 12 valid nodes are retained, and the shape of the valid node data after clipping is 600×12×3.

[0044] In some embodiments, defining the spatial window includes: S111. Taking the tunnel boring machine as the reference point, the ultra-wideband positioning system is used to obtain the real-time coordinates of the reference point.

[0045] UWB positioning tags can be installed on the tunnel boring machine body. The tag refresh frequency is 10 times per second. The UWB base station receives the tag signal and calculates the real-time coordinates of the tunnel boring machine through the arrival time difference algorithm ( , , ).

[0046] S112. Determine the coordinates of the sensor and calculate the actual distance between the sensor and the reference point.

[0047] All sensors in the mine (gas, air volume, temperature sensors) are pre-installed with UWB positioning tags, and their coordinates are obtained through the same UWB system ( , , ).

[0048] S113. Filter out sensors whose actual distances meet the preset distances.

[0049] For each sensor, calculate the Euclidean distance between it and the reference point of the tunnel boring machine. If the spatial window radius is defined as 5 meters, the screening condition is the Euclidean distance. ≤5 meters.

[0050] Calculate the Euclidean distance between all sensors and the tunnel boring machine in real time , generate a valid sensor index list. For example, there are 20 sensors in the mine, and 12 of them meet the requirements in the current window. ≤5 meters, these 12 sensors are screened out.

[0051] The screened sensor data directly reflects the status of high-risk areas around the tunnel boring machine, and can improve the accuracy of subsequent risk index calculations.

[0052] In some embodiments, the dynamic risk index is calculated based on the mine environment parameters, including: 210. For valid node data in the current time window and space window: calculate the mean of gas concentration according to the sampling frequency and spatial node distribution of gas concentration; normalize the real-time air volume, and calculate the mean of normalized real-time air volume according to the sampling frequency and spatial node distribution of real-time air volume; calculate the mean of temperature according to the sampling frequency and spatial node distribution of temperature.

[0053] Take the arithmetic average of the gas concentration values ​​of all sensors in the valid node data to get the mean value of the gas concentration .

[0054] To scale the real-time wind volume value to [0,1], the formula can be used: ;in, and Respectively represent the minimum and maximum values ​​of the real-time wind volume of all sensors in the current space window. Represents the real-time wind volume value before the i-th scaling, Represents the i-th scaled real-time wind volume value.

[0055] Take the arithmetic average of the air volume values ​​of all sensors in the valid node data to get the mean air volume .

[0056] Take the arithmetic average of the temperature values ​​of all sensors in the valid node data to get the mean temperature .

[0057] 220. Based on the mean value of gas concentration, the mean value of normalized real-time air volume and the mean value of temperature, calculate the gas concentration risk factor, air volume risk factor and temperature risk factor.

[0058] The gas concentration risk factor is calculated based on the degree of deviation between the mean gas concentration and the safety threshold: ;in, represents the gas concentration risk factor, Represents the safe threshold of gas concentration.

[0059] The air volume risk factor is calculated based on the deviation between the normalized air volume mean and the ideal value: ;in, Represents the air volume risk factor.

[0060] The temperature risk factor is calculated based on the deviation of the temperature mean from the standard operating temperature range:

[0061] in, Represents the minimum value of the standard operating temperature range, such as 20°C. Represents the maximum value of the standard operating temperature range, such as 35°C.

[0062] 230. Calculate the dynamic risk index based on the gas concentration risk factor, air volume risk factor and temperature risk factor.

[0063] The dynamic risk index is: ;in, represents the dynamic risk index, , , They represent the weights of gas risk factor, air volume risk factor and temperature risk factor respectively.

[0064] The dynamic risk index combines the synergistic effects of gas, air volume, and temperature to improve the accuracy of early warning. , , It can be adjusted dynamically according to the excavation stage.

[0065] For the historical data of each excavation stage, the information entropy of gas concentration risk factor, air volume risk factor and temperature risk factor can be calculated respectively. The weight is determined by calculating the proportion of the information entropy of each factor to the total information entropy. , , .

[0066] In some embodiments, generating a risk trend status code according to a dynamic risk index in S300 includes: Sa310. Determine two or more consecutively increasing risk levels, each risk level corresponds to a unique code; define a demarcation threshold value that is one less than the number of risk levels, and all demarcation threshold values ​​are arranged in ascending order.

[0067] Sa320. Determine the threshold interval of the dynamic risk index and output the risk level code according to the coding mapping mechanism.

[0068] The encoding mapping mechanism includes: Mapping the dynamic risk index below the minimum cut-off threshold to the lowest risk level code; The dynamic risk index between every two adjacent demarcation thresholds is mapped to the intermediate risk level code in sequence according to the order of the demarcation thresholds; Dynamic risk indices that are greater than or equal to the maximum demarcation threshold are mapped to the highest risk level code.

[0069] For example, define three risk levels: low risk (coded 0), medium risk (coded 1), and high risk (coded 2), and two demarcation thresholds (denoted as =0.3 and = 0.6), all the demarcation thresholds are arranged in ascending order as follows: The demarcation threshold is determined by historical data statistics or expert experience. For example, the 50% quantile of the dynamic risk index in historical mine accidents is used as , the 80% quantile is .

[0070] The current dynamic risk index Interval matching with a sequence of demarcation thresholds: like , it is judged as low risk and mapped to code 0.

[0071] like , it is judged as medium risk and mapped to code 1.

[0072] like , it is judged as high risk and mapped to code 2.

[0073] In some embodiments, the adaptive atmospheric pressure model includes an input layer, a first specified number of dilated convolutional layers, and an output layer.

[0074] The input layer receives a series of atmospheric pressures for a specified time range in the past.

[0075] The first specified number of dilated convolutional layers are assigned the optimal dilation rate to capture the changing characteristics of atmospheric pressure under different periods, obtain multiple feature maps, concatenate the feature maps by channels, and integrate multi-scale information into feature tensors.

[0076] The output layer flattens the feature tensor to obtain the time series feature vector.

[0077] The input layer receives the atmospheric pressure sequence within the specified time range in the past, such as the atmospheric pressure data acquired at a sampling frequency of once per second in the past 24 hours, forming a one-dimensional input vector. The dimension of the input data is L×1, where L represents the time step.

[0078] First, a specified number of dilated convolutional layers, for example 3 layers, are assigned a dynamically determined optimal dilation rate, such as the dilation rate =2, =4, = 8. The dilated convolutional layer captures the atmospheric pressure variation patterns at different time scales by adjusting the dilation rate.

[0079] For the k-th layer of dilated convolution, its dilation rate is , the convolution kernel size is K, this layer performs an expansion convolution operation on the input sequence X and outputs a feature map , the formula is: ;in, Represents the weight of the i-th position in the convolution kernel of the k-th layer, and t is the current time step.

[0080] Different dilation rates correspond to different receptive field ranges. For example, a convolutional layer with a dilation rate of d=2 captures short-term fluctuations, such as minute-level changes, while a layer with d=8 captures long-term trends, such as hour-level changes.

[0081] The feature maps output by each dilated convolutional layer are as follows: , , , spliced ​​according to the channel dimension to form a multi-scale feature tensor , so that the model can simultaneously retain the air pressure change information of different time granularities.

[0082] The output layer transforms the feature tensor Flattened into a one-dimensional time series feature vector V, which is used as the input of the subsequent ventilation volume prediction model. The flattening operation can be achieved through a fully connected layer or global pooling.

[0083] By dynamically allocating the optimal expansion rate, the adaptive atmospheric pressure model can automatically select the expansion rate combination according to the periodic characteristics of the current input sequence.

[0084] For example, when high-frequency oscillations in air pressure are detected, the model assigns a smaller expansion rate to the shallow convolution to capture details; when the air pressure changes slowly, the deep convolution uses a large expansion rate to expand the receptive field, significantly improving the ability of the time series feature vector to represent complex air pressure patterns.

[0085] In some embodiments, in the first specified number of dilated convolutional layers being assigned an optimal dilation rate, obtaining the optimal dilation rate includes: Sb310. Get the atmospheric pressure series within the specified time range in the past.

[0086] For example, the data of the past 24 hours with a sampling interval of 1 minute has a total length L=1440.

[0087] Sb320. Perform multi-scale decomposition on the atmospheric pressure sequence to obtain multiple sub-sequences.

[0088] Wavelet transform is used to perform multi-scale decomposition to obtain multiple subsequences.

[0089] For example, using 3-layer discrete wavelet decomposition, we can generate approximate coefficient subsequences (low frequency trend) and detail coefficient subsequences (high frequency noise), (Intermediate frequency fluctuation), (low-frequency period), a total of 4 subsequences, each subsequence has the same length as the original sequence.

[0090] Sb330. Extract the statistical features of the subsequence to obtain a feature vector.

[0091] The following statistical features can be extracted from each subsequence to form a feature vector: Time domain characteristics: mean, variance, skewness, and kurtosis.

[0092] Frequency domain features: The main frequency amplitude is calculated by fast Fourier transform.

[0093] Nonlinear features: approximate entropy.

[0094] Sb340. Cluster the feature vectors of all subsequences to obtain a second specified number of cluster centers and a cluster label to which each subsequence belongs.

[0095] Perform K-means clustering on the feature vectors of all subsequences. The second specified number can be 3 to obtain the cluster centers. , , and the cluster label of each subsequence.

[0096] Sb350. Classify the initial expansion rate for each cluster center and obtain an initial expansion rate set.

[0097] If the subsequence belongs to a high-frequency feature cluster, such as Corresponding , assigning a smaller dilation rate d=2 to capture details.

[0098] If it belongs to the medium frequency cluster, such as Corresponding , assigning a medium expansion rate d=4.

[0099] If it belongs to low-frequency clustering, such as Corresponding , assign a larger equal expansion rate d=8.

[0100] Then the initial expansion rate set is {2, 4, 8}.

[0101] Sb360. When training the adaptive atmospheric pressure model, calculate the importance weight of the subsequence corresponding to each cluster in the current input.

[0102] The importance weight of each cluster subsequence in the current input is calculated through the attention mechanism. For the kth cluster, the weight is .

[0103] Sb370. According to the importance weight of the subsequence corresponding to each cluster in the current input, the initial expansion rate is fine-tuned to obtain the final expansion rate.

[0104] According to weight Initial expansion rate For example, if =0.6 (high frequency features dominate), then the corresponding expansion rate is adjusted to ,in, , represents the adjustment coefficient, and the final expansion rate is obtained .

[0105] Sb380. Construct a set of final expansion rates, use the set of final expansion rates to train the adaptive atmospheric pressure model and make feedback adjustments to obtain the optimal expansion rate.

[0106] The adjusted expansion rate set is input into the temporal convolutional network for forward propagation, the MAE loss function is used to calculate the prediction error, and the model parameters are updated through back propagation.

[0107] At the same time, the cluster centers are dynamically updated according to the performance of the validation set. For example, every 10 training cycles, Sb320-Sb350 is re-executed with the latest data to update the cluster centers and expansion rate allocation rules to form a closed-loop optimization.

[0108] When the tunnel boring machine is blasting, there will be an instantaneous spike in the atmospheric pressure sequence, resulting in a surge in high-frequency components. At this time, the weight of the high-frequency cluster subsequence is dynamically increased. , and lower its expansion rate to d = 1, so that the model convolution kernel can cover a finer time granularity, such as second-level changes, thereby accurately capturing the characteristics of sudden air pressure fluctuations.

[0109] In some embodiments, the ventilation volume prediction interval within a future specified time range is obtained based on the time series feature vector, the risk trend state code, the tunneling state of the tunnel boring machine, and the real-time air volume, including: The improved GBRT architecture is used to process the time series feature vector, risk trend state coding, tunneling status of the tunnel boring machine and real-time air volume, and output the ventilation volume prediction interval within the specified time range in the future.

[0110] Among them, the GBRT improved architecture includes: input layer, feature fusion layer, gradient boosting tree set layer and output layer.

[0111] The input layer is used to receive time series feature vectors, risk trend state encoding, tunneling status of the tunnel boring machine and real-time wind volume.

[0112] The time series feature vector represents the multi-scale time series characteristics of historical air pressure changes; the risk trend state encoding can use 3D unique hot encoding, such as low risk [1,0,0], medium risk [0,1,0], high risk [0,0,1]; The tunneling state of the tunnel boring machine can be a binary parameter, such as 0 represents shutdown and 1 represents tunneling; the real-time air volume represents the real-time air volume value at the current moment.

[0113] All input data are aligned by timestamp to ensure that data within the same time window is processed synchronously.

[0114] The feature fusion layer is used to splice the time series feature vector and the risk trend state encoding vector by channel to form a fusion feature; perform sliding window mean filtering on the real-time air volume to generate air volume trend features; and splice the fusion features, air volume trend features, and tunneling state parameters into a joint feature vector.

[0115] The real-time wind volume sequence is subjected to sliding window mean filtering. For example, the wind volume mean in the window is calculated with a 5-minute window and a 1-minute step size, and the 30-minute data is converted into a 25-dimensional trend feature to characterize the recent wind volume change direction, such as increase, decrease or stability.

[0116] When the excavation status is 1, the corresponding dimension in the joint feature vector is marked as 1, triggering the model to adjust the weights of excavation-related features.

[0117] The gradient boosting tree ensemble layer contains a third specified number of regression trees, each of which performs node splitting based on the joint feature vector and outputs the predicted base value of ventilation volume at each time step.

[0118] The gradient boosting tree set layer contains a specified number (such as 100) of regression trees. Each tree splits the node based on the joint feature vector. When splitting, the information gain of each feature is calculated, and the feature with the largest gain and the threshold are selected to divide the data.

[0119] For example, if the fifth dimension of the wind volume trend feature is greater than 0.8 and the risk code is high risk, the node is split into a left subtree (predicting high wind volume) and a right subtree (predicting low wind volume).

[0120] Each regression tree can predict the base value of ventilation volume at a specified time step in the future (e.g., 12 time steps, one hour in the future, and one step every 5 minutes). The prediction results of 100 trees constitute a base value matrix for subsequent interval calculations.

[0121] The output layer is used to perform quantile regression on the ventilation volume prediction base value, and the quantile loss function is used to calculate the upper and lower quantiles of the future specified time range to generate the ventilation volume prediction interval.

[0122] The output layer uses the quantile loss function to directly predict the upper and lower bounds of the interval. For example, if the upper quantile is set to 0.95 and the lower quantile is set to 0.05, the model minimizes the quantile loss so that 95% of the historical true values ​​fall within the prediction interval.

[0123] The interval width is dynamically scaled based on the ratio of the current dynamic risk index to the historical maximum risk value. For example, when the dynamic risk index reaches 80% of the historical peak, the upper bound quantile is increased from 0.95 to 0.98, the lower bound is reduced from 0.05 to 0.02, and the interval width is expanded by 6% to cope with the uncertainty under high risk.

[0124] When the risk is coded as high risk, the GBRT improved architecture can automatically enhance the sensitivity to sudden changes in air pressure (such as a sudden drop in air pressure caused by an explosion) and adjust the prediction interval in a timely manner.

[0125] When the excavation status is 1, the features related to equipment operation in the regression tree (such as the excavation status is 1 and the air volume trend is greater than 0.5) trigger a high base value prediction, which can reflect the incremental demand for ventilation volume in the excavation operation.

[0126] In some embodiments, the quantile output layer dynamically scales the upper and lower bounds of the ventilation volume prediction interval according to the ratio of the current dynamic risk index to the historical risk interval, and the scaling factor is positively correlated with the dynamic risk index.

[0127] Pre-collect historical data of dynamic risk index of mines during normal operation cycle and calculate its maximum value and minimum value , constituting the historical risk interval[ , ].

[0128] The dynamic risk index calculated in real time Mapped to the historical interval, calculate the relative ratio Ratio, Ratio is: .

[0129] Set zoom factor It is linearly positively correlated with Ratio, setting ,in, Represents a preset constant, such as 0.2.

[0130] The interval expansion ratio required to cover 95% of extreme situations can be determined by analyzing the distribution of prediction errors in historical high-risk events (such as gas concentration exceeding limit alarms).

[0131] Initial upper quantile Can be set to 0.95, the lower bound quantile It can be set to 0.05, and the dynamic scaling rules are: , ;in, and Respectively represent the upper and lower quantiles after dynamic scaling.

[0132] When the upper quantile is dynamically scaled >1 or lower bound quantile after dynamic scaling When <0, it is forced to be truncated to 1.0 or 0.0 to ensure that the quantile is legal.

[0133] When the gas concentration suddenly increases, the dynamic risk index Close to maximum hour, The prediction interval is expanded to cover a wider range of uncertainties. For example, when the gas concentration exceeds the limit alarm, such as the dynamic risk index =0.85, the prediction interval width is expanded from the default 0.9 to 1.0 to improve safety redundancy.

[0134] When the dynamic risk index When it is lower, =0.2, Reduce and narrow the interval width to avoid energy waste caused by excessive ventilation.

[0135] In some embodiments, Figure 2 As shown, the present disclosure provides a ventilation data analysis system for intelligent ventilation in mines, which includes: a multi-source environmental parameter perception module, a dynamic risk assessment and fusion module, a risk coding and air pressure time series feature generation module, and a multi-modal ventilation decision module.

[0136] The multi-source environmental parameter perception module is used to obtain mine environmental parameters, which include gas concentration, real-time air volume and temperature.

[0137] The dynamic risk assessment and fusion module is used to calculate the dynamic risk index based on the mine environment parameters.

[0138] The risk coding and atmospheric pressure time series feature generation module is used to generate risk trend state coding according to the dynamic risk index; obtain the atmospheric pressure sequence within the specified time range in the past, and use the adaptive atmospheric pressure model to process the atmospheric pressure sequence. The adaptive atmospheric pressure model is based on the time convolution network, dynamically determines the optimal expansion rate of the expansion convolution layer, and outputs the time series feature vector after deep integration of multi-scale information; The multimodal ventilation decision module is used to obtain the excavation status of the tunnel boring machine; according to the time series feature vector, risk trend state code, the excavation status of the tunnel boring machine and the real-time air volume, the ventilation volume prediction interval within the specified time range in the future is obtained.

[0139] In some embodiments, the present disclosure provides a device comprising a memory and a processor; the memory is used to store a computer program; the processor is used to implement the steps of a ventilation data analysis method for intelligent ventilation in mines when executing the computer program.

[0140] In some embodiments, the present disclosure provides a readable storage medium, in which computer program instructions are stored. When the computer program instructions are read and executed by a processor, the steps of a ventilation data analysis method for intelligent ventilation in mines are executed.

[0141] Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory.

[0142] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0143] Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent substitutions for some of the technical features therein; and these modifications or substitutions do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A ventilation data analysis method for intelligent ventilation in mines, characterized in that: include: Obtain mine environment parameters, including gas concentration, real-time air volume and temperature; Calculate the dynamic risk index based on the mine environment parameters; Generate risk trend status code based on dynamic risk index; Obtaining an atmospheric pressure sequence within a specified time range in the past, and processing the atmospheric pressure sequence using an adaptive atmospheric pressure model, wherein the adaptive atmospheric pressure model is based on a temporal convolutional network, dynamically determines an optimal dilation rate of a dilated convolutional layer, and outputs a time series feature vector after deep integration of multi-scale information; The tunneling status of the tunnel boring machine is obtained; based on the time series feature vector, risk trend state code, tunneling status of the tunnel boring machine and real-time air volume, the ventilation volume prediction interval within the specified time range in the future is obtained.

2. The ventilation data analysis method for mine intelligent ventilation according to claim 1, characterized in that: Obtain mine environment parameters, including: Configure basic collection parameters, including: setting the sampling frequency of gas concentration, real-time air volume and temperature, defining the time window and space window, as well as the sliding interval of the time window and the sliding interval of the space window; Based on the basic acquisition parameters, a three-dimensional time series tensor including the time axis, spatial node distribution and parameter channels is generated by sliding the time window; According to the sliding of the spatial window, the three-dimensional time series tensor is dynamically clipped to obtain valid node data.

3. The ventilation data analysis method for mine intelligent ventilation according to claim 2, characterized in that: The Define Space window includes: Taking the tunnel boring machine as the reference point, the ultra-wideband positioning system is used to obtain the real-time coordinates of the reference point; Determine the coordinates of the sensor and calculate the actual distance between the sensor and the reference point; Filter out sensors whose actual distances meet the preset distances.

4. The ventilation data analysis method for mine intelligent ventilation according to claim 2, characterized in that: Calculate the dynamic risk index based on the mine environment parameters, including: For valid node data in the current time window and space window: calculate the mean value of gas concentration according to the sampling frequency and spatial node distribution of gas concentration; normalize the real-time air volume, and calculate the mean value of normalized real-time air volume according to the sampling frequency and spatial node distribution of real-time air volume; calculate the mean value of temperature according to the sampling frequency and spatial node distribution of temperature; Calculate the gas concentration risk factor, air volume risk factor and temperature risk factor based on the mean value of gas concentration, the mean value of normalized real-time air volume and the mean value of temperature; The dynamic risk index is calculated based on the gas concentration risk factor, air volume risk factor and temperature risk factor.

5. The ventilation data analysis method for mine intelligent ventilation according to claim 1, characterized in that: Generate risk trend status code based on dynamic risk index, including: Determine two or more risk levels that increase in sequence, each with a unique code; define demarcation thresholds that are one less than the number of risk levels, with all demarcation thresholds arranged in ascending order; Determine the threshold interval of the dynamic risk index and output the risk level code according to the coding mapping mechanism; The encoding mapping mechanism includes: Mapping the dynamic risk index below the minimum cut-off threshold to the lowest risk level code; The dynamic risk index between every two adjacent demarcation thresholds is mapped to the intermediate risk level code in sequence according to the order of the demarcation thresholds; Dynamic risk indices that are greater than or equal to the maximum demarcation threshold are mapped to the highest risk level code.

6. The ventilation data analysis method for mine intelligent ventilation according to claim 1, characterized in that: The adaptive atmospheric pressure model includes an input layer, a first specified number of dilated convolutional layers, and an output layer; The input layer is used to receive the atmospheric pressure series within a specified time range in the past; The first specified number of dilated convolutional layers are assigned the optimal dilation rate to capture the changing characteristics of atmospheric pressure under different periods, obtain multiple feature maps, concatenate the feature maps by channel, and integrate multi-scale information into feature tensors; The output layer flattens the feature tensor to obtain the time series feature vector.

7. The ventilation data analysis method for mine intelligent ventilation according to claim 6, characterized in that: In the first specified number of dilated convolutional layers being assigned the optimal dilation rate, obtaining the optimal dilation rate includes: Get the atmospheric pressure series within the specified time range in the past; Performing multi-scale decomposition on the atmospheric pressure sequence to obtain multiple subsequences; Extract the statistical features of the subsequence to obtain the feature vector; Cluster the feature vectors of all subsequences to obtain the second specified number of cluster centers and the cluster label to which each subsequence belongs; Classify the initial expansion rate for each cluster center to obtain an initial expansion rate set; When training the adaptive atmospheric pressure model, the importance weight of the subsequence corresponding to each cluster in the current input is calculated; According to the importance weight of the subsequence corresponding to each cluster in the current input, the initial expansion rate is fine-tuned to obtain the final expansion rate; A set of final expansion rates is constructed, and the adaptive atmospheric pressure model is trained and feedback-adjusted using the set of final expansion rates to obtain the optimal expansion rate.

8. The ventilation data analysis method for mine intelligent ventilation according to claim 1, characterized in that: According to the time series feature vector, risk trend state code, tunneling state of the tunnel boring machine and real-time air volume, the ventilation volume prediction interval within the specified time range in the future is obtained, including: The GBRT improved architecture is used to process the time series feature vector, risk trend state coding, tunneling state of the tunnel boring machine and real-time air volume, and output the ventilation volume prediction interval within the specified time range in the future; Among them, the GBRT improved architecture includes: input layer, feature fusion layer, gradient boosting tree set layer and output layer; The input layer is used to receive the time series feature vector, risk trend state code, tunneling state of the tunnel boring machine and real-time air volume; The feature fusion layer is used to splice the time series feature vector and the risk trend state encoding vector by channel to form a fusion feature; perform sliding window mean filtering on the real-time air volume to generate air volume trend features; splice the fusion features, air volume trend features, and tunneling state parameters into a joint feature vector; The gradient boosting tree set layer contains a third specified number of regression trees, each of which performs node splitting based on the joint feature vector and outputs the predicted base value of ventilation volume at each time step; The output layer is used to perform quantile regression on the ventilation volume prediction base value, and the quantile loss function is used to calculate the upper and lower quantiles of the future specified time range to generate the ventilation volume prediction interval.

9. The ventilation data analysis method for mine intelligent ventilation according to claim 8, characterized in that: The quantile output layer dynamically scales the upper and lower bounds of the ventilation volume prediction interval according to the ratio of the current dynamic risk index to the historical risk interval, and the scaling coefficient is positively correlated with the dynamic risk index.

10. A ventilation data analysis system for intelligent ventilation in mines, characterized in that: It includes: A multi-source environmental parameter perception module, which is used to obtain mine environment parameters, including gas concentration, real-time air volume and temperature; Dynamic risk assessment and fusion module, which is used to calculate the dynamic risk index based on the mine environment parameters; A risk coding and air pressure time series feature generation module, which is used to generate risk trend state coding according to the dynamic risk index; Obtaining an atmospheric pressure sequence within a specified time range in the past, and processing the atmospheric pressure sequence using an adaptive atmospheric pressure model, wherein the adaptive atmospheric pressure model is based on a temporal convolutional network, dynamically determines an optimal dilation rate of a dilated convolutional layer, and outputs a time series feature vector after deep integration of multi-scale information; The multimodal ventilation decision module is used to obtain the excavation status of the tunnel boring machine; according to the time series feature vector, risk trend state code, the excavation status of the tunnel boring machine and the real-time air volume, the ventilation volume prediction interval within the specified time range in the future is obtained.

Images