A coal mine carbon emission dynamic accounting system based on multi-modal data fusion

Abstract

The present application relates to the field of deep learning carbon accounting, and discloses a coal mine carbon emission dynamic accounting system based on multi-modal data fusion, comprising: a data preprocessing module for collecting multi-source monitoring data and converting it into a multi-dimensional feature vector; a data distribution module for inputting the multi-dimensional feature vector into all polynomial tree dendrite branches; an order adaptive module for making a regression with an adaptive order as a polynomial degree to output a preliminary response value; a gated routing screening module for receiving the preliminary response value, calculating a weight, multiplying the current weight with a historical decay weight to obtain a gated activation value, and screening with a dynamic threshold; a cell summation module for summing the transmitted output to obtain a membrane potential value; a neuron accounting module for mapping the membrane potential value into an accounting value; and a whole mine summary module for weighting and summarizing all accounting values to obtain a total carbon emission accounting value. The present application realizes high-precision dynamic accounting of mine carbon emissions.

Description

Technical Field

[0001] This invention relates to the field of deep learning carbon accounting technology, and in particular to a dynamic carbon emission accounting system for coal mines based on multimodal data fusion. Background Technology

[0002] Coal mining is a significant industrial source of greenhouse gases such as methane. Constructing an accurate and efficient dynamic carbon emission accounting system is a key technological pathway to achieving the "dual carbon" goals (carbon emission control and carbon sequestration). Existing technologies have already explored various approaches to carbon emission accounting in coal mines.

[0003] Regarding accounting methods, existing technologies include methods for measuring greenhouse gas emissions from coal mine exhaust gas and methods for calculating the carbon footprint of coal products. In terms of system architecture, some solutions construct a coal mine carbon emission accounting management system through data acquisition, data processing, and carbon emission accounting modules; some solutions build carbon emission prediction models based on digital twins and LSTM; and some solutions integrate carbon footprint with BIM models to achieve full-cycle carbon emission measurement in mines. Regarding data fusion, existing research has used enterprise ERP systems to acquire production energy consumption and supply chain energy consumption data for automated carbon accounting and intelligent correction; and has constructed multi-source heterogeneous data fusion models to achieve spatiotemporal dynamic monitoring and accurate accounting of methane emissions in the coal industry.

[0004] However, existing technologies still have the following shortcomings. First, the data sources are relatively singular or the fusion level is shallow. Most schemes rely on a single type of data source, and the fusion of multi-source heterogeneous data is mostly limited to the statistical level, failing to fully explore the deep nonlinear correlations between different modalities. Second, existing accounting models mostly use fixed-structure neural networks or traditional regression methods. Their model complexity lacks an adaptive matching mechanism with data distribution characteristics, limiting their generalization ability when faced with significant differences in emission characteristics under different mining areas and operating conditions. Third, in multi-branch network structures, the contribution of each branch to the final accounting result is statically weighted or averaged, lacking a competitive selection mechanism to dynamically adjust branch weights based on input data. This makes it difficult for the model to focus on the most relevant feature channels under complex operating conditions, affecting the accuracy and interpretability of the accounting. Summary of the Invention

[0005] One objective of this invention is to propose a dynamic accounting system for coal mine carbon emissions based on multimodal data fusion. This invention fully integrates multinomial dendritic networks, multimodal data fusion, spatial adaptive order allocation mechanism, and branch competition gating routing mechanism. It describes in detail the entire process of coal mine carbon emissions from multi-source data acquisition, multi-dimensional feature transformation, adaptive regression of dendritic branch order, inter-branch competition screening and historical accumulation gating, cell aggregation accounting, to spatial weighted aggregation of the entire mine. It has the advantages of strong heterogeneous data fusion capability, adaptive matching of model complexity, high interpretability of accounting results, effective suppression of branch noise, and good generalization performance across mining areas.

[0006] According to an embodiment of the present invention, a dynamic accounting system for coal mine carbon emissions based on multimodal data fusion includes: The data preprocessing module is used to collect multi-source monitoring data from coal mines and convert the multi-source monitoring data into multi-dimensional feature vectors. The data distribution module is used to input multidimensional feature vectors into all polynomial dendritic branches; The polynomial dendritic branch order adaptive module is used to perform polynomial regression on multidimensional feature vectors with the learnable space adaptive order as the polynomial degree to output the initial response value. The space adaptive order of each polynomial dendritic branch converges to an independent value during training. The gated routing filtering module is used to receive the preliminary response values ​​of all polynomial dendritic branches within the same neuron, calculate the soft maximum competition weight of each preliminary response value, calculate the historical contribution cumulative value of each soft maximum competition weight over a number of past time points, calculate the current soft maximum competition weight of each polynomial dendritic branch to obtain the gated activation value, set the output of polynomial dendritic branches with gated activation values ​​lower than the dynamic threshold to zero, and pass the output of polynomial dendritic branches with gated activation values ​​not lower than the dynamic threshold to the cell body. The cell body summation module is used to sum the outputs of the polynomial dendritic branches transmitted to the cell body to obtain the membrane potential value. The neuron kernel calculation module is used to map membrane potential values ​​through activation functions and output the calculated kernel values. The whole mine summary module is used to sum up the calculated values ​​output by all neurons according to spatial coordinates to obtain the total carbon emission calculation value of the whole mine.

[0007] Optionally, the data preprocessing module converts multi-source monitoring data into multi-dimensional feature vectors by including the following steps: Connect to methane concentration sensor, wind speed sensor, temperature sensor, ventilation fan negative pressure sensor and mining equipment power metering terminal to collect raw time series data of each monitoring point within a continuous time window; The missing time data was replaced with a weighted sum of the historical average of each sensor and the current measurement value of adjacent points; Replace the current measurement with the median of the measurements taken at the two times before and after the current time. Using the system clock as a reference, data with a sampling frequency lower than the reference is linearly upsampled, and data with a sampling frequency higher than the reference is downsampled at equal intervals; The data is mapped to a zero-mean unit variance distribution using the global mean and standard deviation of each channel; The data from each channel are encoded into multidimensional feature vectors according to sensor type and spatial location.

[0008] Optionally, the multidimensional feature vector input of the data distribution module to all polynomial dendritic branches specifically includes the following steps: Receive the multidimensional feature vector from the data preprocessing module, obtain the total number of branches of the current neuron, copy the multidimensional feature vector into multiple copies with the same number of branches, and send each copy to the input of a corresponding polynomial dendritic branch. Each polynomial dendritic branch receives a copy of a multidimensional feature vector of the same dimension. At the same time, each polynomial dendritic branch uses this copy as input to perform parallel polynomial regression calculations. The calculation processes of each branch are independent of each other and there is no data interaction.

[0009] Optionally, the polynomial dendritic branch order adaptive module performs polynomial regression on the multidimensional feature vector with the learnable space adaptive order as the polynomial degree to output the preliminary response value. Specifically, the module outputs the preliminary response value as follows: Each polynomial dendritic branch receives a multidimensional feature vector from the data distribution module. This multidimensional feature vector contains several dimensions, each of which corresponds to a normalized measurement of a physical quantity at a spatial location. Each polynomial dendritic branch is configured with an independent spatial adaptive order parameter, which is given a random initial value during system initialization and participates in gradient descent updates as a learnable parameter during training. The polynomial dendritic branch uses the spatial adaptive order as the highest power of the polynomial regression. It performs a weighted polynomial expansion on each dimension of the multidimensional feature vector, calculates the first power term, second power term and so on up to the power term of this branch for the measured value of each dimension. Each power term is multiplied by the corresponding trainable weight coefficient and then accumulated. At the same time, a trainable bias term is added to obtain the polynomial regression calculation value of this branch. The polynomial dendritic branch takes the calculated polynomial regression value as the initial response value, performs nonlinear mapping through the branch activation function, and outputs the result. The branch activation function compresses any real number input to the interval between zero and one, and the mapping result is the initial response value output by this branch.

[0010] Optionally, the spatial adaptive order of each polynomial dendritic branch converges to an independent value during training, specifically as follows: During the system training phase, a carbon accounting loss function is constructed based on the labeled true values ​​of carbon emission accounting. A compression excitation term is introduced into the loss function to compress the spatial adaptive order of each polynomial dendritic branch. The compression excitation term applies a compression excitation effect to the spatial adaptive order of each branch in the direction of numerical reduction. The gradient descent optimization algorithm is used to jointly update all trainable parameters. The gradient of each spatial adaptive order parameter consists of two superimposed parts. The first part comes from the gradient signal of the carbon nucleus loss function backpropagated to this branch through the multinomial regression expansion. The second part comes from the gradient signal of the compression excitation term to this branch. The two gradient signals together drive the spatial adaptive order to shift in the direction that reduces the total loss function. There are no shared constraints or coupling constraints on the order parameters between branches. The spatial adaptive order of each branch only receives the gradient signal from the polynomial regression part within its own branch and the gradient signal from the compression excitation term to this branch, and does not receive the gradient signal from other branches. Training is terminated when the number of training iterations reaches a preset upper limit or the change in the total loss function is lower than a preset convergence threshold. The spatial adaptive order of each branch converges to a different independent value at the end of training because the gradient sources are independent and the gradient evolution paths are different.

[0011] Optionally, the gated routing filtering module receives the preliminary response values ​​of all polynomial dendritic branches within the same neuron and calculates the soft maximum competition weight for each preliminary response value as follows: The gated routing filtering module obtains the preliminary response values ​​output by all polynomial dendritic branches within the current neuron. There are several branches within the same neuron, and each branch outputs a preliminary response value. All preliminary response values ​​constitute a set of real numbers. The gated routing filtering module performs soft maximum contention weight calculation on all preliminary response values, and uses the soft maximum contention weight of each branch as the contention score of that branch at the current moment.

[0012] Optionally, the gated routing filtering module sets the output of polynomial dendritic branches with gated activation values ​​lower than the dynamic threshold to zero and transmits the output of polynomial dendritic branches with gated activation values ​​not lower than the dynamic threshold to the cell body as follows: The gated routing filtering module obtains the current soft maximum competition weight of each polynomial dendritic branch, obtains the historical contribution accumulation value of each branch output by the historical accumulation module, calculates the current soft maximum competition weight of each branch and the corresponding historical contribution accumulation value, and uses the calculation result as the gate activation value of this branch. The gated routing filtering module calculates the average value of all polynomial dendritic branch gating activation values ​​at the current moment and uses it as a dynamic threshold.

[0013] The gated routing filtering module compares the gated activation value of each branch with the dynamic threshold. Branches with gated activation values ​​lower than the dynamic threshold are identified as low-contribution branches, and the initial response value output by this branch is set to zero, thus blocking the transmission of the branch signal to the cell body. The gated routing filtering module determines branches with gated activation values ​​not lower than the dynamic threshold as high-contribution branches, preserves the original value of the initial response value output by this branch and passes it to the cell body, allowing this branch signal to participate in the cell body summation operation; The gated routing filtering module recalculates the soft maximum competition weight, gated activation value, and dynamic threshold of all branches at each moment, so that the activation state of each branch changes dynamically with the input data.

[0014] Optionally, the cell body summation module sums the outputs of the polynomial dendritic branches transmitted to the cell body to obtain the membrane potential value, specifically including the following steps: Obtain the initial response values ​​of all polynomial dendritic branches passed to the cell body by the gated routing filtering module. Each initial response value is the original value output by the branch in the gated routing filtering module whose gated activation value is not lower than the dynamic threshold. All the initial response values ​​obtained are aggregated inside the cell body, and the resulting value is used as the total input value of the neuron's cell body. The total input value of the cell body is output as the membrane potential value of the neuron to the neuron nucleus calculation module.

[0015] Optionally, the neuron kernel calculation module's process of mapping membrane potential values ​​through an activation function and then outputting the calculated kernel value specifically includes the following steps: The neuron nucleus calculation module receives the membrane potential value output by the cell body summation module; The neuron kernel module uses the membrane potential value as the input variable of the activation function. The activation function is a monotonically increasing function that compresses the input variable to the interval between zero and one. After function mapping, it outputs a kernel value between zero and one. The neuron kernel module outputs the calculated value as the carbon emission intensity representation value of the spatial location corresponding to the current input data to the whole mine summary module.

[0016] Optionally, the process of summarizing the calculated values ​​output by all neurons in the whole mine according to spatial coordinates to obtain the total carbon emission calculation value of the whole mine specifically includes the following steps: Receive all the calculated values ​​output by the neuron kernel modules, with each calculated value accompanied by the downhole spatial coordinate label corresponding to this neuron; Based on the spatial coordinate labels of each neuron, a mapping relationship between the calculated values ​​and spatial positions is established in the three-dimensional spatial coordinate system, with each spatial coordinate position corresponding to a calculated value. Assign a spatial weight coefficient to each spatial coordinate position; The calculated value for each spatial coordinate location is multiplied by the corresponding spatial weight coefficient to obtain the weighted calculated value for each spatial location. The weighted accounting values ​​of all spatial locations are aggregated and summarized in the spatial dimension, and the summarized value is output as the total carbon emission accounting value of the entire mine.

[0017] The beneficial effects of this invention are: (1) By independently configuring a learnable spatial adaptive order for each polynomial dendritic branch, the model can automatically match the optimal polynomial degree according to the nonlinear characteristics of data under different mining areas and working conditions. The order of each branch converges to an independent value during training, realizing a refined complexity allocation of one branch and one order. This solves the problem of underfitting or overfitting of traditional fixed-order polynomial regression under complex data distribution, and significantly improves the accuracy of carbon emission accounting for the entire mine. (2) By introducing a soft maximum competition mechanism and time decay history accumulation between branches through the gating route filtering module, the dendritic branches within the same neuron compete dynamically according to the input data. Only the branches with gating activation values ​​not lower than the dynamic threshold are output to the cell body, effectively suppressing noise interference from low-contribution branches. This allows the model to maintain stable accounting performance under downhole conditions such as sensor data fluctuations and communication packet loss. The explicit expression of multinomial regression gives the accounting results good interpretability, making it easy for field engineers to trace the accounting basis. (3) The gated routing filtering module recalculates the soft maximum competition weight, gated activation value and dynamic threshold of all branches at each time, so that the activation state of each branch changes dynamically with the input data. During the inference stage, only the high contribution branches are activated to participate in cell summation, which greatly reduces the computational cost of a single calculation and adapts to the computing power constraints at the edge of the coal mine. The dual mechanism of order adaptation and branch competition enables the model to have good cross-scenario generalization ability for different geological conditions of mining areas and emission characteristics of different production stages. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a dynamic accounting system for coal mine carbon emissions based on multimodal data fusion proposed in this invention. Figure 2 This is a schematic diagram of the polynomial order space adaptive allocation mechanism of a coal mine carbon emission dynamic accounting system based on multimodal data fusion proposed in this invention.

[0019] Figure 3This is a schematic diagram of the inter-branch competition gating routing mechanism of a coal mine carbon emission dynamic accounting system based on multimodal data fusion proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figures 1-3 A dynamic accounting system for coal mine carbon emissions based on multimodal data fusion, comprising: The data preprocessing module is used to collect multi-source monitoring data from coal mines and convert the multi-source monitoring data into multi-dimensional feature vectors. The data distribution module is used to input multidimensional feature vectors into all polynomial dendritic branches; The polynomial dendritic branch order adaptive module is used to perform polynomial regression on multidimensional feature vectors with the learnable space adaptive order as the polynomial degree to output the initial response value. The space adaptive order of each polynomial dendritic branch converges to an independent value during training. The gated routing filtering module is used to receive the preliminary response values ​​of all polynomial dendritic branches within the same neuron, calculate the soft maximum competition weight of each preliminary response value, multiply each soft maximum competition weight over several past time points by a time decay coefficient and then sum them to obtain the historical contribution accumulation value, multiply the current soft maximum competition weight of each polynomial dendritic branch by the historical contribution accumulation value to obtain the gated activation value, set the output of polynomial dendritic branches with gated activation values ​​lower than the dynamic threshold to zero, and pass the output of polynomial dendritic branches with gated activation values ​​not lower than the dynamic threshold to the cell body; The cell body summation module is used to sum the outputs of the polynomial dendritic branches transmitted to the cell body to obtain the membrane potential value. The neuron kernel calculation module is used to map membrane potential values ​​through activation functions and output the calculated kernel values. The whole mine summary module is used to sum up the calculated values ​​output by all neurons according to spatial coordinates to obtain the total carbon emission calculation value of the whole mine.

[0022] In this embodiment, the data preprocessing module converts multi-source monitoring data into multi-dimensional feature vectors, including the following steps: Connect to methane concentration sensor, wind speed sensor, temperature sensor, ventilation fan negative pressure sensor and mining equipment power metering terminal to collect raw time series data of each monitoring point within a continuous time window; The missing time data was replaced with a weighted sum of the historical average of each sensor and the current measurement value of adjacent points; Replace the current measurement with the median of the measurements taken at the two times before and after the current time. Using the system clock as a reference, data with a sampling frequency lower than the reference is linearly upsampled, and data with a sampling frequency higher than the reference is downsampled at equal intervals; The data is mapped to a zero-mean unit variance distribution using the global mean and standard deviation of each channel; The data from each channel are encoded into multidimensional feature vectors according to sensor type and spatial location.

[0023] Encoding the data from each channel into multidimensional feature vectors according to sensor type and spatial location specifically includes the following steps: Obtain the category labels and spatial coordinate labels of all sensor channels. The category labels include methane concentration, wind speed, temperature, negative pressure and power. The spatial coordinate labels are the position coordinates of each sensor in the three-dimensional coordinate system of the mine. The three axes of the three-dimensional coordinate system are along the mine strike direction, dip direction and vertical direction respectively. Using sensor category as the primary classification criterion, all channels are divided into multiple category groups, with each category group containing channel data for all spatial locations within that category. Using spatial coordinates as the second classification criterion, the channel data are arranged in coordinate order within each category group. The coordinate order is arranged sequentially from the intake air side to the return air side along the mine direction. Channel data of different categories at the same coordinate position are assigned to their respective category groups. The standardized values ​​of each channel are concatenated dimension by dimension according to the order of the category group and the coordinate order within the group. The standardized value of each channel occupies one dimension of the feature vector. After all channels are concatenated, a one-dimensional vector is formed, which is the multi-dimensional feature vector.

[0024] In this embodiment, the process of inputting the multidimensional feature vector of the data distribution module into all polynomial dendritic branches specifically includes the following steps: Receive the multidimensional feature vector from the data preprocessing module, obtain the total number of branches of the current neuron, copy the multidimensional feature vector into multiple copies with the same number of branches, and send each copy to the input of a corresponding polynomial dendritic branch. Each polynomial dendritic branch receives a copy of a multidimensional feature vector of the same dimension. At the same time, each polynomial dendritic branch uses this copy as input to perform parallel polynomial regression calculations. The calculation processes of each branch are independent of each other and there is no data interaction.

[0025] Specifically, the parallel polynomial regression calculation performed by each polynomial dendritic branch at the same time using this copy as input includes the following steps: The data distribution module simultaneously sends the same copy of the multidimensional feature vector to the input of all polynomial dendritic branches, and each polynomial dendritic branch receives the copy within the same system clock cycle. Each polynomial dendritic branch extracts the values ​​of all dimensions from the copy, and each branch independently obtains its current spatial adaptive order parameter as well as its stored polynomial weight coefficients and bias terms. Each polynomial dendritic branch uses the values ​​of each dimension of the copy as independent variables and its own spatial adaptive order as the highest power of the polynomial. The power values ​​of each dimension's independent variables are calculated from the first power to the power of the order. Each power value is multiplied by the corresponding weight coefficient and then accumulated. A bias term is then added to obtain the polynomial regression calculation result of that branch. Each polynomial dendritic branch performs independent multiplication and addition operations during the above calculation process. There is no data exchange or dependency between the operation paths of each branch. The weight coefficients and bias terms of each branch are stored independently in the registers of their respective branches.

[0026] Specifically, each polynomial dendritic branch obtains its current spatial adaptive order parameter as follows: Each polynomial dendritic branch is assigned a random initial order value during initialization. This initial order value is generated by uniform random sampling when the system starts up, and the sampling range is a continuous interval between a preset minimum order and a preset maximum order. Each polynomial dendritic branch has its own independent storage space in the system memory, which is used to store the order parameter, polynomial weight coefficients and bias term of the branch; During the forward computation phase of each training round, each polynomial dendritic branch reads the currently saved order parameter value from its own independent storage space as the highest power of the polynomial for this regression calculation. During the backpropagation phase of each training round, the system calculates the gradient of the order parameter of all branches according to the loss function. The order parameter of each branch only receives the gradient signal from the backpropagation of the polynomial regression calculation path within that branch and the gradient signal from the compression activation term acting on that branch alone. The order parameters of each branch are updated independently based on the gradient signals they receive, and the updated order parameters are written back to the branch's own independent storage space to overwrite the old values. In the forward computation phase of the next training round, each branch reads the updated order parameter values ​​from its own storage space, and this process is repeated iteratively until training terminates.

[0027] The order parameter storage spaces of each branch are isolated from each other, and there are no cross-branch order parameter read or write operations.

[0028] The sampling range is a continuous interval between a preset minimum order and a preset maximum order. The preset minimum order and maximum order are set as follows: The minimum order is preset to 1. When the order is 1, the polynomial regression degenerates into linear regression, which can only fit the linear mapping relationship between the input and the output. Setting the minimum order to 1 ensures that each dendritic branch has at least linear fitting ability, and avoids the branch output being constant due to the order dropping to zero, thus losing the ability to respond to the input variables. The preset maximum order is determined based on the total number of dimensions of the input feature vectors of the polynomial dendritic branch. The total number of dimensions of the multidimensional feature vectors received by the branch is obtained and used as the benchmark value of the preset maximum order. When the total number of dimensions is less than 4, the preset maximum order is fixed at 4. When the total number of dimensions is greater than or equal to 4, the preset maximum order is equal to the integer part of the sum of the total number of dimensions and the sum of the total number of dimensions minus one multiplied by 0.5. The upper limit of the preset maximum order is fixed at 6. When the value calculated in the above way exceeds 6, the preset maximum order is set to 6. This upper limit is determined based on the statistical distribution that the total number of feature dimensions of multi-source monitoring data in coal mines is usually between 10 and 42. Limiting the maximum order to within 6 can ensure the expressive power of the model while controlling the total number of parameters of each branch within the allowable range of storage capacity and computing bandwidth of edge devices. The preset minimum order of 1 and the preset maximum order of 6 together constitute the numerical constraint boundary of the spatial adaptive order parameter during the training process.

[0029] In this embodiment, the polynomial dendritic branch order adaptive module performs polynomial regression on the multidimensional feature vector with the learnable space adaptive order as the polynomial degree to output the preliminary response value. Specifically, the module outputs the preliminary response value as follows: Each polynomial dendritic branch receives a multidimensional feature vector from the data distribution module. This multidimensional feature vector contains several dimensions, each of which corresponds to a normalized measurement of a physical quantity at a spatial location. Each polynomial dendritic branch is configured with an independent spatial adaptive order parameter, which is given a random initial value during system initialization and participates in gradient descent updates as a learnable parameter during training. The polynomial dendritic branch uses the spatial adaptive order as the highest power of the polynomial regression. It performs a weighted polynomial expansion on each dimension of the multidimensional feature vector, calculates the first power term, second power term and so on up to the power term of this branch for the measured value of each dimension. Each power term is multiplied by the corresponding trainable weight coefficient and then accumulated. At the same time, a trainable bias term is added to obtain the polynomial regression calculation value of this branch. The polynomial dendritic branch takes the calculated polynomial regression value as the initial response value, performs nonlinear mapping through the branch activation function, and outputs the result. The branch activation function compresses any real number input to the interval between zero and one, and the mapping result is the initial response value output by this branch.

[0030] The weighted polynomial expansion of each dimension of the multidimensional feature vector specifically includes the following steps: Obtain the multidimensional feature vector received by the polynomial dendritic branch at the current time step. The multidimensional feature vector contains multiple dimensions, and each dimension corresponds to a normalized measurement value. Obtain the currently saved spatial adaptive order parameter of the dendritic branch of the polynomial, and take the integer part of the order parameter as the highest power of the polynomial expansion. For each dimension in the multidimensional feature vector, calculate the power value of the measured value of that dimension from the first power to the second power up to the highest power. The power values ​​of each dimension constitute the power value sequence of that dimension. Obtain the polynomial weight coefficients stored in the polynomial dendritic branch, which contain multiple independent coefficients that correspond one-to-one with each power of each dimension. Multiply each power value of each dimension by its corresponding weight coefficient to obtain the weighted result of each power for each dimension; The sum of all power-weighted results across all dimensions is then added, followed by the bias term stored in the dendritic branch of the polynomial. The resulting value is the polynomial regression calculation value for that branch.

[0031] The output is obtained by performing a non-linear mapping using a branch activation function, where the activation function is the sigmoid function.

[0032] The process involves nonlinear mapping using a branch activation function, which compresses any real input to the range of zero to one. The mapping result is the initial response value of this branch output. The specific steps include: Obtain the calculated value of the multinomial regression and use this value as the input variable for the branch activation function; The branch activation function uses the Sigmoid function, which is based on the mapping relationship that the output variable is equal to one divided by the sum of one and the negative power of the natural constant e of the input variable; Substituting the calculated value of the polynomial regression into the above mapping relationship, the corresponding output variable value is calculated, which is between zero and one. The value of this output variable is used as the initial response value of the polynomial dendritic branch at the current moment and output to the gated routing filtering module.

[0033] In this embodiment, the spatial adaptive order of each polynomial dendritic branch converges to an independent value during training as follows: During the system training phase, a carbon accounting loss function is constructed based on the labeled true carbon emission accounting values. This loss function uses the sum of squared deviations between the total carbon emission accounting values ​​and the true values ​​of the entire mine as the optimization objective. A compression excitation term is introduced into the loss function to compress the spatial adaptive order of each polynomial dendritic branch. The compression excitation term applies a compression excitation effect to the spatial adaptive order of each branch in the direction of numerical reduction. The gradient descent optimization algorithm is used to jointly update all trainable parameters. The gradient of each spatial adaptive order parameter consists of two superimposed parts. The first part comes from the gradient signal of the carbon nucleus loss function backpropagated to this branch through the multinomial regression expansion. The second part comes from the gradient signal of the compression excitation term to this branch. The two gradient signals together drive the spatial adaptive order to shift in the direction that reduces the total loss function. There are no shared constraints or coupling constraints on the order parameters between branches. The spatial adaptive order of each branch only receives the gradient signal from the polynomial regression part within its own branch and the gradient signal from the compression excitation term to this branch, and does not receive the gradient signal from other branches. Training is terminated when the number of training iterations reaches a preset upper limit or the change in the total loss function is lower than a preset convergence threshold. The spatial adaptive order of each branch converges to a different independent value at the end of training because the gradient sources are independent and the gradient evolution paths are different.

[0034] Specifically, the compression excitation term is: In the backpropagation calculation, the system calculates the partial derivative of each order parameter in the expression. For the order parameter of a single branch, the derivative of its absolute value function is positive 1 when the parameter is greater than zero, negative 1 when the parameter is less than zero, and zero when the parameter is equal to zero using the subgradient. Multiplying the partial derivative by the preset regularization coefficient, the resulting product is the additional gradient signal received by the branch order parameter. The direction of the additional gradient signal always points to zero. That is, when the parameter is positive, the additional gradient is positive, causing the parameter to be updated in the decreasing direction. When the parameter is negative, the additional gradient is negative, causing the parameter to be updated in the increasing direction. The additional gradient signal is added to the principal gradient signal obtained by differentiating the carbon loss function with respect to the same order parameter in each training step, and together they determine the update direction and step size of the order parameter. The magnitude of the additional gradient signal is controlled by a preset regularization coefficient. The larger the coefficient value, the larger the absolute value of the additional gradient signal, and the stronger the compression of the order parameter towards zero. The calculation process is performed independently on each branch, and the additional gradient signal of each branch is determined only by the order parameter value of that branch itself, without involving data interaction across branches.

[0035] The specific method for setting the preset regularization coefficient here is as follows: The baseline value is determined by the ratio of the logarithm of the total number of training samples to the preset maximum order value. The baseline value is then multiplied by a preset scaling factor to obtain the initial value of the regularization coefficient. During system training, this coefficient remains constant and is not updated with each training round; the value range of this coefficient is set between 0.001 and 0.1. When it is detected that the spatial adaptive order of more than 80% of the polynomial dendritic branches converges to the preset minimum order in multiple consecutive training rounds, it indicates that the compression intensity is too high. The current regularization coefficient is reduced to half of the original value and training is restarted. When it is detected that the spatial adaptive order of more than 50% of the polynomial dendritic branches is still equal to the preset maximum order after training, it indicates that the compression intensity is insufficient. The current regularization coefficient is increased to twice the original value and training is restarted.

[0036] The joint update of all trainable parameters using the gradient descent optimization algorithm specifically includes the following steps: Obtain the carbon accounting loss function value and the compression excitation term value for the current training batch, and add the two together to obtain the total loss function value; Perform backpropagation calculation on the total loss function value to obtain the spatial adaptive order parameter of the total loss function with respect to each polynomial dendritic branch, the weight coefficients corresponding to each order, and the partial derivatives of the bias terms. The update step size of each parameter is obtained by multiplying the partial derivative of each parameter by the preset learning rate. The preset learning rate is set to the range of 0.001 to 0.01 during system initialization and gradually decreases exponentially with the increase of training rounds. Subtract the update step size corresponding to each parameter from the current value of that parameter to obtain the new value of that parameter after the update; All trainable parameters are updated synchronously within the same training step. All the new values ​​are written to the storage space corresponding to each parameter, replacing the old values ​​before the update, and then proceed to the next training round.

[0037] The gradient of each spatial adaptive order parameter consists of two superimposed parts, specifically including the following steps: The carbon accounting loss function is backpropagated to obtain the gradients of the weight coefficients and bias terms in the polynomial regression expansion of each branch. The gradients of each weight coefficient and bias term are backpropagated along the power connection path in the polynomial regression expansion to the spatial adaptive order parameter of that branch. The gradient values ​​transmitted on each power connection path are summed according to the path, and the result is used as the first part of the gradient. The partial derivative of the compression excitation term with respect to the spatial adaptive order parameter of a single branch is calculated. When the order parameter value of the branch is greater than zero, the partial derivative is positive 1; when the order parameter value of the branch is less than zero, the partial derivative is negative 1; and when the order parameter value of the branch is equal to zero, the partial derivative is zero. The partial derivative is multiplied by the preset regularization coefficient, and the resulting product is used as the second part of the gradient. The first part of the gradient is added to the second part of the gradient, and the sum is used as the total gradient value of the adaptive order parameter of the branch space. The total gradient value acts independently on the order parameter of each branch. The order parameter of each branch only receives the first and second part of the gradient corresponding to its own branch, and does not receive the gradient signal of other branches.

[0038] In this embodiment, the gated routing filtering module receives the preliminary response values ​​of all polynomial dendritic branches within the same neuron and calculates the soft maximum competition weight for each preliminary response value as follows: The gated routing filtering module obtains the preliminary response values ​​output by all polynomial dendritic branches within the current neuron. There are several branches within the same neuron, and each branch outputs a preliminary response value. All preliminary response values ​​constitute a set of real numbers. The gated routing filtering module performs soft maximum contention weight calculation on all preliminary response values, and uses the soft maximum contention weight of each branch as the contention score of that branch at the current moment.

[0039] Specifically, each branch outputs an initial response value as follows: After each polynomial dendritic branch completes the weighted polynomial expansion of the multidimensional feature vector, the polynomial regression value of that branch is obtained, which is a real number. The branch inputs the polynomial regression calculation value into the branch activation function. The branch activation function, the Sigmoid function, compresses the input variable to the interval between zero and one and outputs a mapping value, which is a real number between zero and one. The branch will use the mapped value output by the branch activation function as the initial response value of the branch at the current moment; The initial response value is simultaneously output to the gated routing filter module and the aggregation node of other branches within the neuron. The gated routing filter module receives the initial response values ​​of all polynomial dendritic branches within the same neuron at the same time.

[0040] The gated routing filtering module performs soft maximum contention weight calculation on all preliminary response values, specifically including the following steps: The gated routing filtering module obtains the preliminary response values ​​of all polynomial dendritic branches within the same neuron at the current time. Each branch outputs a preliminary response value, and the preliminary response values ​​of all branches constitute a set of real numbers. The gated routing filtering module calculates the natural exponent value for each initial response value in the set of real numbers, and obtains the exponent value corresponding to each branch. The natural exponent value is raised to the power of the initial response value of the branch with the natural constant as the base. The gated routing filtering module sums all natural index values ​​to obtain the sum of all natural index values; The gated routing filtering module divides the natural index value of each branch by the sum of all natural index values, and the resulting ratio is the soft maximum competition weight of that branch. The gated routing filtering module outputs the soft maximum contention weight of all branches to the historical cumulative calculation unit and the gated activation calculation unit inside the gated routing filtering module.

[0041] In this embodiment, the gated routing filtering module sets the output of polynomial dendritic branches with gated activation values ​​lower than the dynamic threshold to zero and transmits the output of polynomial dendritic branches with gated activation values ​​not lower than the dynamic threshold to the cell body as follows: Obtain the current soft maximum competition weight of each polynomial dendritic branch, obtain the historical contribution accumulation value of each branch output by the historical accumulation module, multiply the current soft maximum competition weight of each branch by the corresponding historical contribution accumulation value, and use the result as the gating activation value of this branch. The average value of all polynomial dendritic branch gating activation values ​​at the current moment is multiplied by a preset scaling factor to obtain the dynamic threshold. The preset scaling factor is set as follows: it is set as a fixed constant during system initialization, with a value range between 0.3 and 0.8, and the initial default value is set to 0.5. This factor is used to multiply the average value of all gating activation values ​​at the current moment, and the product is used as the dynamic threshold. During the training process, this factor remains constant and does not change with the training rounds. After the training is completed, the factor is fine-tuned according to the calculation accuracy index on the validation set: the factor value that achieves the highest calculation accuracy is searched in the range of 0.3 to 0.8 with a step size of 0.1, and the factor value is selected as the fixed factor for the final deployment.

[0042] The gating activation value of each branch is compared with the dynamic threshold. Branches with gating activation values ​​lower than the dynamic threshold are identified as low-contribution branches. The initial response value of the branch is set to zero, thus blocking the transmission of the branch signal to the cell body. Branches with gating activation values ​​not lower than the dynamic threshold are identified as high-contribution branches. The initial response value output by this branch is retained in its original form and passed to the cell body, allowing the branch signal to participate in the cell body summation operation. At each moment, the soft maximum competition weight, gating activation value and dynamic threshold of all branches are recalculated, so that the activation state of each branch switches dynamically as the input data changes.

[0043] In this embodiment, the cell body summation module sums the outputs of the polynomial dendritic branches transmitted to the cell body to obtain the membrane potential value, specifically including the following steps: Obtain the initial response values ​​of all polynomial dendritic branches passed to the cell body by the gated routing filtering module. Each initial response value is the original value output by the branch in the gated routing filtering module whose gated activation value is not lower than the dynamic threshold. All the initial response values ​​obtained are aggregated inside the cell body, and the resulting value is used as the total input value of the neuron's cell body. The total input value of the cell body is output as the membrane potential value of the neuron to the neuron nucleus calculation module.

[0044] The specific steps involved in pooling all the acquired preliminary response values ​​within the cell body are as follows: The cell body summation module receives all preliminary response values ​​passed to the cell body by the gated routing filtering module. Each preliminary response value is the original value output by the branch in the gated routing filtering module whose gated activation value is not lower than the dynamic threshold. Each preliminary response value is accompanied by the neuron index number to which the branch belongs. The cell body summation module performs an accumulation and aggregation operation on all received preliminary response values ​​within the cell body, summing all preliminary response values ​​under the same neuron index number to obtain the total cell body input value of that neuron; The cell body summation module takes the total cell body input value obtained after addition and summation as the membrane potential value, and outputs the membrane potential value to the neuron nucleus calculation module.

[0045] The process of outputting the total cell body input value as the neuron's membrane potential value to the neuron's nucleus calculation module specifically includes the following steps: The cell body summation module assigns the total cell body input value obtained after addition to the membrane potential variable of the neuron. The membrane potential variable occupies a fixed storage space in the internal storage unit of the cell body summation module. After the cell body summation module completes the assignment of the membrane potential variable, it sends the membrane potential variable from the output port of the cell body summation module to the input port of the neuron nucleus calculation module. During the transmission process, the membrane potential variable is accompanied by the neuron index number as identification information. After receiving the membrane potential variable, the neuron nucleus calculation module returns an acknowledgment signal to the cell body summation module. Upon receiving the acknowledgment signal, the cell body summation module clears the temporary storage area of ​​the membrane potential variable corresponding to the neuron at the current moment, and prepares for the calculation and output of the membrane potential value at the next moment.

[0046] In this embodiment, the neuron kernel calculation module outputs the calculated kernel value after mapping the membrane potential value through the activation function, specifically including the following steps: The neuron nucleus calculation module receives the membrane potential value output by the cell body summation module; The neuron kernel module uses the membrane potential value as the input variable of the activation function. The activation function is a monotonically increasing function that compresses the input variable to the interval between zero and one. After function mapping, it outputs a kernel value between zero and one. The neuron kernel module outputs the calculated value as the carbon emission intensity representation value of the spatial location corresponding to the current input data to the whole mine summary module.

[0047] The specific steps involved in outputting a calculated value between zero and one after function mapping are as follows: The neuron nucleus calculation module receives the membrane potential value output by the cell body summation module and uses this membrane potential value as the input variable of the activation function. This membrane potential value is a real number. The neuron kernel calculation module calls the activation function calculation unit stored inside the module. The activation function calculation unit has a pre-set mapping rule for the Sigmoid function. The mapping rule is that the output variable is equal to one divided by the sum of one and the negative power of the natural constant e of the input variable. The neuron kernel calculation module substitutes the membrane potential value into the mapping rule to calculate and obtain a real output value between zero and one. The magnitude of this output value reflects the carbon emission intensity of the spatial location corresponding to the neuron. A value close to one represents a high emission intensity, and a value close to zero represents a low emission intensity. The neuron's core calculation module uses the real number output value as the core calculation value output by the neuron and sends it to the mine-wide summary module.

[0048] Specifically, the neuron kernel calculation module outputs the calculated value as the carbon emission intensity representation value of the spatial location corresponding to the current input data to the whole mine summary module, which includes the following steps: The neuron kernel module reads the preset spatial coordinate label of the neuron from the internal storage unit. The spatial coordinate label is a three-dimensional coordinate value, which corresponds to the position in the mine strike direction, dip direction and vertical direction respectively. The neuron kernel module binds the calculated values ​​with spatial coordinate labels to form a data pair. In the data pair, the calculated values ​​come first and the spatial coordinate labels come second. The calculated values ​​and spatial coordinate labels are concatenated in a fixed-length format. The neuron kernel module sends the data pair to the input port of the mine-wide summary module through the module output port. The sending process is carried out in the form of data packets, and the header of the data packet contains the neuron index number and the total length of the data packet. After receiving the data packet, the whole mine summary module parses out the calculation values ​​and spatial coordinate labels, and returns a reception confirmation signal to the neuron calculation module; After receiving the confirmation signal, the neuron kernel module clears the kernel values ​​and spatial coordinate labels that have been sent in the temporary storage area corresponding to the neuron at the current moment, and prepares for the data output at the next moment.

[0049] In this embodiment, the total carbon emission calculation value of the entire mine is obtained by weighting and summing the calculation values ​​output by all neurons according to spatial coordinates in the whole mine summary module. The specific steps include the following: Receive all the calculated values ​​output by the neuron kernel modules, with each calculated value accompanied by the downhole spatial coordinate label corresponding to this neuron; Based on the spatial coordinate labels of each neuron, a mapping relationship between the calculated values ​​and spatial positions is established in the three-dimensional spatial coordinate system, with each spatial coordinate position corresponding to a calculated value. Assign a spatial weight coefficient to each spatial coordinate position; The calculated value for each spatial coordinate location is multiplied by the corresponding spatial weight coefficient to obtain the weighted calculated value for each spatial location. The weighted accounting values ​​of all spatial locations are aggregated and summarized in the spatial dimension, and the summarized value is output as the total carbon emission accounting value of the entire mine.

[0050] The process of aggregating and summarizing the weighted values ​​of all spatial locations along the spatial dimension includes the following steps: The whole mine summary module reads the weighted accounting values ​​corresponding to all spatial coordinate positions from the internal storage unit, and each weighted accounting value is accompanied by a three-dimensional spatial coordinate label of that position; The whole mine summary module uses three-dimensional spatial coordinate labels as indexes to obtain the total number of all spatial coordinate locations; The entire mine summary module starts from the first spatial coordinate position and sequentially adds the weighted accounting value of that position to the cumulative variable. After adding the weighted accounting value of each position, it moves to the next position until the weighted accounting value of all spatial coordinate positions has been added. The total coal mine summary module uses the accumulated variable value after the summation as the total carbon emission accounting value for the entire coal mine. The mine-wide summary module outputs the total carbon emissions of the entire mine to the system output port, which is connected to an external display terminal or data storage medium.

[0051] Example 1: This invention is applied to a dynamic carbon emission accounting scenario at a coal mine working face. The working face employs longwall mining, producing approximately 3000 tons of raw coal per day. Intake and return airways are arranged along the strike, with a total length of approximately 2200 meters from the mining face to the return air outlet. Monitoring points are deployed at key locations such as the return airway, the working face, the gas drainage pump station, and the main electromechanical chambers. Each monitoring point integrates a methane concentration sensor, a wind speed sensor, a temperature sensor, and a negative pressure sensor. Power metering terminals are also installed on major electrical equipment such as the coal mining machine, scraper conveyor, and local ventilation fans. All sensors continuously collect data at a sampling frequency of 1 Hz and transmit it to a ground-based edge computing server via an industrial Ethernet ring network. The server deploys the accounting system proposed in this invention, which includes a data preprocessing module, a data distribution module, a polynomial dendritic branch order adaptive module, a gating routing filtering module, a cell body summation module, a neuron accounting module, and a mine-wide aggregation module.

[0052] In this scenario, the system operates as follows: The data preprocessing module first receives the raw time-series data from all acquisition terminals. For each channel, missing values ​​are imputed using a weighted sum of the channel's mean from the previous time period and the current measurement from spatially adjacent channels. Abnormal jump values ​​are replaced with the median of the two sampling points before and after the current time. Then, all channel data are unified to the same timeline using the system clock as a reference. Finally, standardization is performed based on the historical statistical values ​​of each channel, outputting a multi-dimensional feature vector. Each dimension represents a normalized value for methane concentration, wind speed, temperature, negative pressure, or equipment power at a specific spatial location. The data distribution module copies this feature vector multiple times and synchronously sends it to each polynomial dendritic branch in the neuron array.

[0053] Each neuron is configured with 8 polynomial dendritic branches. After receiving the feature vector, the j-th branch of the i-th neuron determines the highest power of the polynomial by the order adaptive module based on the spatial adaptive order parameter k_ij currently held by the branch. After multiple rounds of training, the k values ​​of each branch converge to 3, 1, 2, 4, 1, 2, 3, 1 respectively. Branch 1 expands each dimension of the feature vector with a 3rd-order polynomial, branch 2 performs linear regression with a 1st-order polynomial, and branch 4 captures highly nonlinear relationships with a 4th-order polynomial. Each branch outputs a preliminary response value. The gating routing filtering module collects the preliminary response values ​​of the 8 branches within the same neuron. First, it calculates the soft maximum competition weight of each branch. For example, the weights obtained after exponential normalization of the preliminary response values ​​of the 8 branches at the current time are 0.05, 0.27, 0.14, 0.31, 0.18, 0.03, 0.02, and 0.00 respectively. The first 4 branches have higher weights, and the last 4 branches have lower weights. The module then accumulates the soft maximum weights of each branch over the past 20 time steps with a decay coefficient of 0.9, yielding a historical cumulative contribution value. The current weight is multiplied by this historical cumulative value to obtain the gating activation value. The gating activation values ​​for each branch are 0.04, 0.25, 0.11, 0.28, 0.01, 0.02, 0.01, and 0.00, respectively. The dynamic threshold is set to the mean of all activation values, 0.09. Branches 1 to 4 have activation values ​​higher than the threshold, and their outputs are passed to the cell body. Branches 5 to 8 have activation values ​​lower than the threshold, and their outputs are set to zero. The cell body summation module performs a convergence operation on the preliminary response values ​​of the four high-contribution branches, outputting the membrane potential value of the neuron. The neuron kernel calculation module maps and compresses the membrane potential value to the 0-1 range using a Sigmoid function, serving as a representation of the carbon emission intensity at the spatial location corresponding to the neuron. The whole mine summary module receives the calculated values ​​of all neurons. Based on the spatial coordinates associated with each neuron, it multiplies the corresponding position by the spatial volume weight of the mining area represented by that coordinate and then performs a global summary, finally outputting the total carbon emission calculation value of the whole mine at the current sampling time.

[0054] To further verify the practical effect of the present invention, the system was deployed at the aforementioned working face and operated continuously for 30 days, collecting approximately 2.59 million valid records. Continuous measurements from the same return air outlet using a high-precision online gas chromatograph were used as a reference benchmark to compare the accuracy of the traditional fixed-order polynomial regression method with that of the present invention. The fixed-order method uniformly adopted a second-order polynomial, while other network structures and training rounds remained consistent with the present invention. Data from four typical monitoring points—the entrance section of the return airway, the middle section of the return airway, the middle of the mining face, and the gas drainage pumping station—were selected, and the results are shown in Table 1.

[0055] Table 1: Comparison of Carbon Emission Accounting Results at Typical Monitoring Sites

[0056] As shown in Table 1, the calculated values ​​of this invention at all four monitoring points are closer to the gas chromatography reference values. The methane concentration at the entrance section of the return airway is only 0.12, belonging to the low-concentration area. The traditional method's calculated value of 126.7 is 7.4 lower than the reference value of 134.1, while the calculated value of this invention, 132.5, is only 1.6 lower than the reference value. After the concentration in the middle section of the return airway rises to 0.28, the deviation of the traditional method begins to widen; 289.4 is 22.4 lower than the reference value of 311.8, while the value of this invention, 306.2, is only 5.6 lower. The concentration in the middle of the mining face reaches 0.89. The traditional method's 936.5 is 70.7 lower than the reference value of 1007.2, while the value of this invention, 998.4, is only 8.8 lower. The highest gas concentration at the gas drainage pumping station reached 1.52. The traditional method's concentration of 1592.7 was 105.6 lower than the reference value of 1698.3, while the present invention's concentration of 1685.2 was only 13.1 lower. The relative deviations of the four points using the traditional method were -5.52, -7.18, -7.02, and -6.22, with an average deviation of approximately -6.49 points. The relative deviations of the four points using the present invention were -1.19, -1.80, -0.87, and -0.77, with an average deviation of approximately -1.16 points, improving the calculation accuracy by approximately 5.33 percentage points. Traditional methods consistently produce negative biases, indicating that the fixed second-order polynomial is insufficient in fitting nonlinear relationships across different concentration ranges, systematically underestimating emissions. This invention, through an adaptive order mechanism, increases the order of the branch corresponding to the middle of the mining face to fourth order and the branch corresponding to the gas drainage pump station to third order, thus capturing more complex mapping relationships. Simultaneously, the branch order is maintained at first order in low-concentration areas to prevent overfitting, reducing the bias from 6-7 points to less than 1 point. The gated routing module assigns a weight of 0.31 to the fourth branch when inputting data from the middle of the mining face and increases the weight of the seventh branch to 0.28 when inputting data from the gas drainage pump station. This mechanism of automatically switching high-contribution branches based on concentration distribution ensures stable model performance across different concentration ranges, avoiding being hampered by linear relationships in low-concentration areas or struggling with strong nonlinear relationships in high-concentration areas, resulting in overall superior performance compared to traditional methods.

[0057] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A dynamic accounting system for coal mine carbon emissions based on multimodal data fusion, characterized in that, include: The data preprocessing module is used to collect multi-source monitoring data from coal mines and convert the multi-source monitoring data into multi-dimensional feature vectors. The data distribution module is used to input multidimensional feature vectors into all polynomial dendritic branches; The polynomial dendritic branch order adaptive module is used to perform polynomial regression on multidimensional feature vectors with the learnable space adaptive order as the polynomial degree to output the initial response value. The space adaptive order of each polynomial dendritic branch converges to an independent value during training. The gated routing filtering module is used to receive the initial response values ​​of all polynomial dendritic branches within the same neuron, calculate the soft maximum competition weight of each initial response value, and calculate the historical contribution cumulative value by calculating each soft maximum competition weight over several past time points. The current soft maximum competition weight of each polynomial dendritic branch is calculated to obtain the gating activation value. The output of polynomial dendritic branches with gating activation values ​​lower than the dynamic threshold is set to zero, and the output of polynomial dendritic branches with gating activation values ​​not lower than the dynamic threshold is passed to the cell body. The cell body summation module is used to sum the outputs of the polynomial dendritic branches transmitted to the cell body to obtain the membrane potential value. The neuron kernel calculation module is used to map membrane potential values ​​through activation functions and output the calculated kernel values. The whole mine summary module is used to sum up the calculated values ​​output by all neurons according to spatial coordinates to obtain the total carbon emission calculation value of the whole mine.

2. The coal mine carbon emission dynamic accounting system based on multimodal data fusion according to claim 1, characterized in that, The data preprocessing module converts multi-source monitoring data into multi-dimensional feature vectors, including the following steps: Connect to methane concentration sensor, wind speed sensor, temperature sensor, ventilation fan negative pressure sensor and mining equipment power metering terminal to collect raw time series data of each monitoring point within a continuous time window; The missing time data was replaced with a weighted sum of the historical average of each sensor and the current measurement value of adjacent points; Replace the current measurement with the median of the measurements taken at the two times before and after the current time. Using the system clock as a reference, data with a sampling frequency lower than the reference is linearly upsampled, and data with a sampling frequency higher than the reference is downsampled at equal intervals; The data is mapped to a zero-mean unit variance distribution using the global mean and standard deviation of each channel; The data from each channel are encoded into multidimensional feature vectors according to sensor type and spatial location.

3. The dynamic accounting system for coal mine carbon emissions based on multimodal data fusion according to claim 1, characterized in that, The data distribution module's multidimensional feature vector input to all polynomial dendritic branches specifically includes the following steps: Receive the multidimensional feature vector from the data preprocessing module, obtain the total number of branches of the current neuron, copy the multidimensional feature vector into several copies with the same number of branches, and send each copy to the input of a corresponding polynomial dendritic branch. Each polynomial dendritic branch receives a copy of a multidimensional feature vector of the same dimension. At the same time, each polynomial dendritic branch uses this multidimensional feature vector copy as input to perform parallel polynomial regression calculations. The calculation processes of each branch are independent of each other and there is no data interaction.

4. The coal mine carbon emission dynamic accounting system based on multimodal data fusion according to claim 1, characterized in that, The polynomial dendritic branch order adaptive module uses the learnable space adaptive order as the polynomial degree to perform polynomial regression on the multidimensional feature vector and outputs the preliminary response value as follows: Each polynomial dendritic branch receives a multidimensional feature vector from the data distribution module. This multidimensional feature vector contains several dimensions, each of which corresponds to a normalized measurement of a physical quantity at a spatial location. Each polynomial dendritic branch is configured with an independent spatial adaptive order parameter, which is given a random initial value during system initialization and participates in gradient descent updates as a learnable parameter during training. The polynomial dendritic branch uses the spatial adaptive order as the highest power of the polynomial regression. It performs a weighted polynomial expansion on each dimension of the multidimensional feature vector, calculates the first power term, second power term and so on up to the power term of this branch for the measured value of each dimension. Each power term is multiplied by the corresponding trainable weight coefficient and then accumulated. At the same time, a trainable bias term is added to obtain the polynomial regression calculation value of this branch. The polynomial dendritic branch takes the calculated polynomial regression value as the initial response value, performs nonlinear mapping through the branch activation function, and outputs the result. The branch activation function compresses any real number input to the interval between zero and one, and the mapping result is the initial response value output by this branch.

5. A dynamic accounting system for coal mine carbon emissions based on multimodal data fusion according to claim 1, characterized in that, The spatial adaptive order of each polynomial dendritic branch converges to an independent value during training, specifically as follows: During the system training phase, a carbon accounting loss function is constructed based on the labeled true values ​​of carbon emission accounting. A compression excitation term is introduced into the loss function to compress the spatial adaptive order of each polynomial dendritic branch. The compression excitation term applies a compression excitation effect to the spatial adaptive order of each branch in the direction of numerical reduction. The gradient descent optimization algorithm is used to jointly update all trainable parameters. The gradient of each spatial adaptive order parameter consists of two superimposed parts. The first part comes from the gradient signal of the carbon nucleus loss function backpropagated to this branch through the multinomial regression expansion. The second part comes from the gradient signal of the compression excitation term to this branch. The two gradient signals together drive the spatial adaptive order to shift in the direction that reduces the total loss function. There are no shared constraints or coupling constraints on the order parameters between branches. The spatial adaptive order of each branch only receives the gradient signal from the polynomial regression part within its own branch and the gradient signal from the compression excitation term to this branch, and does not receive the gradient signal from other branches. Training is terminated when the number of training iterations reaches a preset upper limit or the change in the total loss function is lower than a preset convergence threshold. At the end of training, each value converges to a different independent value.

6. The coal mine carbon emission dynamic accounting system based on multimodal data fusion according to claim 1, characterized in that, The gated routing filtering module receives the preliminary response values ​​of all polynomial dendritic branches within the same neuron and calculates the soft maximum competition weight for each preliminary response value as follows: The gated routing filtering module obtains the preliminary response values ​​output by all polynomial dendritic branches within the current neuron. There are several branches within the same neuron, and each branch outputs a preliminary response value. All preliminary response values ​​constitute a set of real numbers. The gated routing filtering module performs soft maximum contention weight calculation on all preliminary response values, and uses the soft maximum contention weight of each branch as the contention score of that branch at the current moment.

7. A dynamic accounting system for coal mine carbon emissions based on multimodal data fusion as described in claim 1, characterized in that, The gated routing filtering module sets the output of polynomial dendritic branches with gated activation values ​​below the dynamic threshold to zero and transmits the output of polynomial dendritic branches with gated activation values ​​not below the dynamic threshold to the cell body as follows: Obtain the current soft maximum competition weight of each polynomial dendritic branch, obtain the historical contribution accumulation value of each branch output by the historical accumulation module, and calculate the average value of the gating activation values ​​of all polynomial dendritic branches at the current moment as the dynamic threshold by combining the current soft maximum competition weight of each branch with the corresponding historical contribution accumulation value. The gating activation value of each branch is compared with the dynamic threshold. Branches with gating activation values ​​lower than the dynamic threshold are identified as low-contribution branches. The initial response value of the branch is set to zero, thus blocking the transmission of the branch signal to the cell body. Branches with gating activation values ​​not lower than the dynamic threshold are identified as high-contribution branches. The initial response value output by this branch is retained in its original form and passed to the cell body, allowing the branch signal to participate in the cell body summation operation. At each moment, the soft maximum competition weight, gating activation value and dynamic threshold of all branches are recalculated, so that the activation state of each branch switches dynamically as the input data changes.

8. The coal mine carbon emission dynamic accounting system based on multimodal data fusion according to claim 1, characterized in that, The cell body summation module sums the outputs of the polynomial dendritic branches transmitted to the cell body to obtain the membrane potential value, specifically including the following steps: Obtain the initial response values ​​of all polynomial dendritic branches passed to the cell body by the gated routing filtering module. Each initial response value is the original value output by the branch in the gated routing filtering module whose gated activation value is not lower than the dynamic threshold. All the initial response values ​​obtained are aggregated inside the cell body, and the resulting value is used as the total input value of the neuron's cell body. The total input value of the cell body is output as the membrane potential value of the neuron to the neuron nucleus calculation module.

9. A dynamic accounting system for coal mine carbon emissions based on multimodal data fusion according to claim 1, characterized in that, The neuron kernel calculation module maps membrane potential values ​​through an activation function and outputs the calculated kernel value, specifically including the following steps: The neuron nucleus calculation module receives the membrane potential value output by the cell body summation module; The neuron kernel module uses the membrane potential value as the input variable of the activation function. The activation function is a monotonically increasing function that compresses the input variable to the interval between zero and one. After function mapping, it outputs a kernel value between zero and one. The neuron kernel module outputs the calculated value as the carbon emission intensity representation value of the spatial location corresponding to the current input data to the whole mine summary module.

10. A dynamic accounting system for coal mine carbon emissions based on multimodal data fusion according to claim 1, characterized in that, The entire mine aggregation module weights and aggregates the calculated values ​​output by all neurons according to spatial coordinates to obtain the total carbon emission calculation value of the entire mine. The specific steps include the following: Receive all the calculated values ​​output by the neuron kernel modules, with each calculated value accompanied by the downhole spatial coordinate label corresponding to this neuron; Based on the spatial coordinate labels of each neuron, a mapping relationship between the calculated values ​​and spatial positions is established in the three-dimensional spatial coordinate system, with each spatial coordinate position corresponding to a calculated value. Assign a spatial weight coefficient to each spatial coordinate position; The calculated value for each spatial coordinate location is multiplied by the corresponding spatial weight coefficient to obtain the weighted calculated value for each spatial location. The weighted accounting values ​​of all spatial locations are aggregated and summarized in the spatial dimension, and the summarized value is output as the total carbon emission accounting value of the entire mine.

Images