A coal spontaneous combustion temperature prediction method based on ant colony algorithm optimized neural network

Abstract

The application discloses a coal spontaneous combustion temperature prediction method based on an ant colony algorithm optimized neural network and belongs to the technical field of coal safety. A current environment state set containing coal body temperature, oxygen concentration and other parameters is acquired; an ant colony algorithm optimization program is called; an iteration process of path exploration and evaluation of an ant colony in a discretized parameter space according to pheromone and heuristic information is simulated; the optimal connection weight and threshold parameter set of the neural network model are automatically searched and determined; the neural network model is instantiated by using the optimized parameter set, and environment state parameters are input; the instantiated model is run for calculation; the coal spontaneous combustion temperature prediction value at a future preset time point is output through nonlinear transformation and feature fusion of the hidden layer of the model. The application overcomes the defects of experience-dependent parameter tuning of a traditional model by using an intelligent optimization algorithm, improves prediction accuracy and model generalization ability, and can adapt to early warning requirements of various scenes such as underground, open air and transportation.

Description

Technical Field

[0001] This application relates to the field of coal safety technology, and in particular to a method for predicting the spontaneous combustion temperature of coal based on an ant colony algorithm-optimized neural network. Background Technology

[0002] Currently, the prediction of coal spontaneous combustion temperature largely relies on classical theoretical models based on physicochemical reactions or traditional machine learning models based on historical statistical data. Theoretical models often require numerous simplifications and assumptions regarding the complex coal-oxygen complex reaction process, making it difficult to accurately depict the actual conditions in variable environments such as underground goafs and open-pit coal yards. On the other hand, traditional machine learning models, such as standard neural networks, heavily depend on initial parameter settings and a large number of high-quality training samples for their predictive performance. When faced with different coal characteristics and various environmental disturbances, the models often exhibit insufficient generalization ability and robustness, and the parameter tuning process relies heavily on empirical trial and error, resulting in low efficiency.

[0003] Existing technologies lack an intelligent prediction method capable of adaptively optimizing core parameters and effectively integrating key influencing factors across different scenarios. This results in prediction accuracy and real-time performance for complex situations such as hidden fire zones in coal mines, dynamic heat sources in open-pit storage yards, and unstable temperature rises during transportation failing to meet the urgent safety warning needs of the modern coal industry. Therefore, there is an urgent need to develop a new method for predicting coal spontaneous combustion temperature that can overcome the aforementioned deficiencies, automatically acquire optimal model parameters through an efficient intelligent search mechanism, and flexibly adapt to various actual working conditions. Summary of the Invention

[0004] To achieve the above objectives, this application provides the following technical solution:

[0005] According to a first aspect of the present invention, the present invention claims a method for predicting the spontaneous combustion temperature of coal based on an ant colony algorithm-optimized neural network, comprising:

[0006] S1: Obtain the current environmental state parameter set of the target coal sample, which includes the internal temperature of the coal body, the ambient oxygen concentration, the coal body porosity, the ambient humidity, and the temperature of the surrounding rock where the coal body is located.

[0007] S2: Call the ant colony algorithm optimization program to optimize and search the initial connection weight matrix and initial neuron threshold vector of the pre-configured neural network model to generate an optimized neural network parameter set;

[0008] S3: Load the optimized neural network parameter set into the corresponding structure of the neural network model to complete the instantiation of the neural network model;

[0009] S4: Input the current environment state parameter set obtained in S1 into the instantiated neural network model;

[0010] S5: Run the instantiated neural network model to perform calculations. The hidden layer of the neural network model performs nonlinear transformation and feature fusion on the input parameters. The output layer calculates the predicted value of the spontaneous combustion temperature of the target coal sample at a future preset time point based on the output result of the hidden layer and outputs the predicted value of the spontaneous combustion temperature.

[0011] Further, in S2, the optimization search includes: constructing a search space that represents candidate neural network parameter solutions using individual ant paths; calculating the state transition probability of ants based on pheromone concentration and heuristic information on the path to explore new solutions; evaluating the quality of ant paths based on the fitting accuracy of the historical coal spontaneous combustion dataset after the candidate neural network parameter solutions are input into the neural network model; dynamically updating the pheromone concentration on the path based on the evaluation result; guiding the ant colony to gather in a region of better parameter solutions; and after multiple iterations, determining the parameter solution corresponding to the globally optimal path as the optimized neural network parameter set.

[0012] The process of evaluating the fitting accuracy of the ant path to the historical coal spontaneous combustion dataset based on the parameter solutions of the candidate neural network inputted into the neural network model, and dynamically updating the pheromone concentration on the path based on the evaluation results, specifically includes:

[0013] S21: For each ant's complete path constructed in the current iteration, decode the path into complete candidate neural network parameters, including the connection weights between all layers and the thresholds of each neuron;

[0014] S22: Assign the candidate neural network parameters to a feedforward neural network with a fixed structure, and perform forward propagation calculation on the network using an independent historical coal spontaneous combustion verification dataset. The historical coal spontaneous combustion verification dataset contains multiple sets of historical environmental state parameters and their corresponding, subsequently observed coal spontaneous combustion temperatures.

[0015] S23: Calculate the mean square error between the predicted temperature and the actual observed temperature of all samples in the historical coal spontaneous combustion verification dataset by the network, and use the reciprocal of the mean square error as the fitness value for evaluating the quality of the current ant path.

[0016] S24: After all ants have completed path construction in one iteration, find the local optimum with the highest fitness value in this iteration and the global optimum with the highest fitness value in all iterations so far;

[0017] S25: Perform a global evaporation operation on all paths of pheromones to simulate the natural evaporation of pheromones;

[0018] S26: Ants corresponding to the local optimal path and the global optimal path are only allowed to add pheromones to the path segments they traverse, and the intensity of the added pheromone is proportional to the fitness value of the corresponding path.

[0019] Furthermore, in step S2, the step of calling the ant colony optimization program to optimize the initial connection weight matrix and initial neuron threshold vector of the pre-configured neural network model to generate an optimized neural network parameter set is specifically implemented through the following detailed steps:

[0020] S2a: Parameter Encoding and Search Space Discretization: All connection weights and neuron thresholds to be optimized in the neural network model are arranged in order to form a multidimensional optimization vector; a discrete and finite value search interval is defined for each parameter dimension in the multidimensional optimization vector, and the interval is evenly divided into N levels, where N is an integer greater than 1; the complete path of each ant is composed of the sequence of level numbers it selects in each parameter dimension, and the path uniquely corresponds to the specific candidate neural network parameter value;

[0021] S2b: Ant colony initialization and pheromone matrix construction: Initialize an ant colony of size M, where M is an integer greater than 1; initialize pheromone concentration values ​​for each of the N levels in each parameter dimension of the multidimensional optimization vector, thereby forming a multidimensional pheromone concentration matrix; at the same time, calculate a static heuristic pheromone value for each level, which is proportional to the probability that the parameter value represented by that level is within a preset reasonable prior range;

[0022] S2c: Iterative search main loop: Set the maximum number of iterations K, where K is an integer greater than 1, and enter the iteration loop;

[0023] S2d: Single Ant Path Construction: For each ant in the colony, select a level for each dimension in the order of the parameter dimensions. For the current dimension, the ant calculates the probability of selecting each level using a roulette wheel selection strategy based on the current pheromone concentration and heuristic information value of all N levels in that dimension. The level is randomly selected according to the calculated probability distribution and serves as the ant's path node in the current dimension. This process is repeated until all dimensions have been selected, thus forming a complete path for the ant.

[0024] S2e: Path evaluation and fitness calculation: For all paths constructed by M ants in the current iteration, decode them into candidate parameters and calculate their corresponding fitness values ​​according to the methods described in S21 to S23.

[0025] S2f: Pheromone Update: Following the methods described in S24 to S26, update the pheromone concentration matrix based on the path evaluation results of the current iteration;

[0026] S2g: Iteration Termination Judgment and Optimal Solution Output: When the number of iterations reaches K, or the fitness value of the global optimal path does not improve significantly in several consecutive iterations, the iteration loop is terminated; the global optimal path in the record is decoded into specific neural network connection weight values ​​and neuron thresholds, and this set of parameters is the optimized neural network parameter set.

[0027] Furthermore, in S3, S4, and S5, the construction, instantiation, and prediction calculation of the neural network model specifically include:

[0028] S3a: Neural Network Structure Definition: The neural network model is a feedforward neural network structure containing an input layer, at least a hidden layer, and an output layer; the number of neurons in the input layer is strictly equal to the number of parameters in the current environmental state parameter set, and each neuron receives the corresponding state parameter value; the output layer contains neurons for outputting the predicted value of the spontaneous combustion temperature; the neurons between each layer are fully connected.

[0029] S3b: Model parameter instantiation: The values ​​in the optimized neural network parameter set generated in S2 are assigned to the corresponding connection weights and neuron thresholds in the neural network model according to the predefined mapping rules; wherein, the connection weight values ​​are assigned to the weight coefficients of each directed edge connecting adjacent layer neurons, and the neuron thresholds are assigned to the internal bias terms of each neuron except the input layer.

[0030] S3c: Forward Propagation Calculation Process: When the current environment state parameter set is input to the input layer, the data begins forward propagation; for each neuron in the hidden layer and output layer, the calculation process is to calculate the weighted sum of the signal values ​​transmitted from all input connections of the neuron and the corresponding connection weights, add this weighted sum to the threshold of the neuron to obtain the net input value of the neuron, and map the net input value through a preset, non-linear activation function to obtain the output value of the neuron; the output value of the previous layer is used as the input signal of the next layer, and is passed layer by layer;

[0031] S3d: Prediction value generation: The final output value of the output layer neuron, after being processed by a linear or preset scaling function, is determined as the predicted value of the spontaneous combustion temperature of the target coal sample at a preset time point in the future; the preset time point is one or more future time intervals pre-set according to the response time requirements of early warning of coal spontaneous combustion.

[0032] Furthermore, the method also includes an adaptive treatment step to address the differences in spontaneous combustion tendency of different coal types:

[0033] Before executing S1, obtain the coal type identification information of the target coal sample;

[0034] Based on the coal type identification information, a corresponding target parameter configuration group is selected from a plurality of pre-stored parameter configuration groups; wherein, different parameter configuration groups are associated with different coal types and at least include configuration parameters related to the ant colony algorithm optimization search in S2, the configuration parameters including the number of ants M, the number of iterations K, the pheromone volatile factor, and the search range of each parameter in the multidimensional optimization vector;

[0035] Use the parameters in the target parameter configuration group to initialize or adjust the running configuration of the ant colony algorithm optimization program, and then execute steps S2 to S6.

[0036] Furthermore, the method also includes specific application steps for underground goaf scenarios in coal mines:

[0037] The set of current environmental state parameters obtained in S1 additionally includes parameters of air leakage intensity in the goaf and parameters of the distribution characteristics of residual coal thickness in the goaf.

[0038] In S5, the hidden layer of the neural network model is configured to have at least two sub-hidden layers, wherein the first sub-hidden layer is specifically used to fuse the parameters of the internal temperature of the coal body, the ambient oxygen concentration, and the air leakage intensity of the goaf to calculate the comprehensive oxidation heating potential index; the second sub-hidden layer receives the comprehensive oxidation heating potential index and other environmental state parameters to perform the final feature integration.

[0039] The output predicted auto-ignition temperature is marked as corresponding to a specific goaf zone or coordinate location.

[0040] Furthermore, the method also includes specific application steps for open-air coal storage yard scenarios:

[0041] The set of current environmental state parameters obtained in S1 additionally includes environmental wind speed and direction, solar radiation intensity, and coal pile surface temperature gradient parameters.

[0042] The method further includes a correction step after S5: based on the current seasonal information and historical meteorological statistics, a dynamic correction offset is applied to the initial spontaneous combustion temperature prediction value calculated by the neural network model. The correction offset is obtained by querying a predefined seasonal-meteorological correction table.

[0043] Furthermore, the method also includes specific application steps for coal transportation scenarios:

[0044] The set of current environmental state parameters obtained in S1 is collected in real time by a sensor network deployed on the transport vehicle, and additionally includes parameters such as the change rate of coal vibration compaction and the local gas residence time in the carriage.

[0045] In S5, the neural network model is periodically invoked in a rolling time window manner. Each time it is invoked, the latest set of environmental state parameters is input, and the predicted curve of spontaneous combustion temperature during the future transportation period is output.

[0046] The predicted auto-ignition temperature value is output together with the geographical location information or transportation mileage information.

[0047] Furthermore, the output of the predicted auto-ignition temperature value in step S6 specifically includes at least one of the following output formats:

[0048] Format 1: The predicted spontaneous combustion temperature is sent in digital form to the graphical user interface for real-time display, and a visual alarm is triggered when the predicted value exceeds a first threshold, and an auditory alarm is triggered when it exceeds a higher second threshold.

[0049] Format 2: The predicted spontaneous combustion temperature value, its corresponding target coal sample identifier, prediction time point, and data acquisition timestamp are encapsulated into a structured data message and sent to the upper-level monitoring system through an industrial communication interface;

[0050] Form 3: Compare the predicted value of the spontaneous combustion temperature with the historical predicted value and the actual observed value to generate a prediction reliability assessment index, and store the index and the predicted value together in a local log file.

[0051] This invention discloses a method for predicting coal spontaneous combustion temperature based on an ant colony optimization neural network, belonging to the field of coal safety technology. It obtains a current environmental state set containing multiple parameters such as coal temperature and oxygen concentration; calls an ant colony optimization program, simulating the iterative process of an ant colony exploring and evaluating paths in a discrete parameter space based on pheromones and heuristic information, automatically searching and determining the optimal connection weights and threshold parameter set for the neural network model; instantiates the neural network model using this optimized parameter set, inputting the environmental state parameters; runs the instantiated model for calculation, and outputs a predicted value of coal spontaneous combustion temperature at a preset future time point through nonlinear transformation and feature fusion of its hidden layers. This invention overcomes the shortcomings of traditional model parameter tuning relying on experience through intelligent optimization algorithms, improves prediction accuracy and model generalization ability, and can adapt to the early warning needs of various scenarios such as underground, open-pit, and transportation. Attached Figure Description

[0052] Figure 1 A flowchart illustrating the workflow of a method for predicting the spontaneous combustion temperature of coal based on an ant colony algorithm optimized neural network, as claimed in an embodiment of the present invention.

[0053] Figure 2 The second flowchart of a method for predicting the spontaneous combustion temperature of coal based on an ant colony algorithm optimized neural network, as claimed in an embodiment of the present invention, is shown below.

[0054] Figure 3 The third flowchart is a method for predicting the spontaneous combustion temperature of coal based on an ant colony algorithm optimized neural network, as claimed in an embodiment of the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0056] The terms "first," "second," and "third" in this application are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications in the embodiments of this application, such as up, down, left, right, front, back, etc., are only used to explain the relative positional relationships and movements between components in a specific orientation as shown in the accompanying drawings. If the specific orientation changes, the directional indications will change accordingly. Furthermore, the terms "including" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0057] References to embodiments herein mean that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0058] According to the first embodiment of the present invention, referring to Figure 1 This invention claims protection for a method for predicting the spontaneous combustion temperature of coal based on an ant colony algorithm-optimized neural network, comprising:

[0059] S1: Obtain the current environmental state parameter set of the target coal sample, which includes the internal temperature of the coal body, the ambient oxygen concentration, the coal body porosity, the ambient humidity, and the temperature of the surrounding rock where the coal body is located.

[0060] S2: Call the ant colony algorithm optimization program to optimize and search the initial connection weight matrix and initial neuron threshold vector of the pre-configured neural network model to generate an optimized neural network parameter set;

[0061] S3: Load the optimized neural network parameter set into the corresponding structure of the neural network model to complete the instantiation of the neural network model;

[0062] S4: Input the current environment state parameter set obtained in S1 into the instantiated neural network model;

[0063] S5: Run the instantiated neural network model to perform calculations. The hidden layer of the neural network model performs nonlinear transformation and feature fusion on the input parameters. The output layer calculates the predicted value of the spontaneous combustion temperature of the target coal sample at a future preset time point based on the output result of the hidden layer and outputs the predicted value of the spontaneous combustion temperature.

[0064] Further, in S2, the optimization search includes: constructing a search space that represents candidate neural network parameter solutions using individual ant paths; calculating the state transition probability of ants based on pheromone concentration and heuristic information on the path to explore new solutions; evaluating the quality of ant paths based on the fitting accuracy of the historical coal spontaneous combustion dataset after the candidate neural network parameter solutions are input into the neural network model; dynamically updating the pheromone concentration on the path based on the evaluation result; guiding the ant colony to gather in a region of better parameter solutions; and after multiple iterations, determining the parameter solution corresponding to the globally optimal path as the optimized neural network parameter set.

[0065] The method involves evaluating the fitting accuracy of the ant path to the historical coal spontaneous combustion dataset after inputting the candidate neural network parameter solutions into the neural network model, and dynamically updating the pheromone concentration on the path based on the evaluation results. Figure 2 Specifically, it includes:

[0066] S21: For each ant's complete path constructed in the current iteration, decode the path into complete candidate neural network parameters, including the connection weights between all layers and the thresholds of each neuron;

[0067] S22: Assign the candidate neural network parameters to a feedforward neural network with a fixed structure, and perform forward propagation calculation on the network using an independent historical coal spontaneous combustion verification dataset. The historical coal spontaneous combustion verification dataset contains multiple sets of historical environmental state parameters and their corresponding, subsequently observed coal spontaneous combustion temperatures.

[0068] S23: Calculate the mean square error between the predicted temperature and the actual observed temperature of all samples in the historical coal spontaneous combustion verification dataset by the network, and use the reciprocal of the mean square error as the fitness value for evaluating the quality of the current ant path.

[0069] S24: After all ants have completed path construction in one iteration, find the local optimum with the highest fitness value in this iteration and the global optimum with the highest fitness value in all iterations so far;

[0070] S25: Perform a global evaporation operation on all paths of pheromones to simulate the natural evaporation of pheromones;

[0071] S26: Ants corresponding to the local optimal path and the global optimal path are only allowed to add pheromones to the path segments they traverse, and the intensity of the added pheromone is proportional to the fitness value of the corresponding path.

[0072] In this embodiment, S21: Path Decoding Operation: In one iteration, a complete path successfully constructed by each ant is represented as a sequence of selection identifiers. The system decodes this path sequence according to a preset rule that is completely opposite to the encoding rule in step S2a. The decoding process is as follows: Based on the selection identifier at each position in the path sequence, it maps back to its corresponding parameter dimension and the level selected in that dimension. Then, based on the median or a preset representative value of the parameter value range represented by that level, the specific floating-point value is restored. All positions in the path sequence are processed sequentially, ultimately restoring a complete set of candidate parameter values ​​that corresponds one-to-one with the parameters to be optimized in the neural network model, including all weight values ​​and all thresholds.

[0073] S22: Temporary Network Configuration and Verification Calculation: The system creates a temporary network computation graph with the exact same number of layers and neurons per layer as the pre-configured neural network model. The candidate parameter values ​​decoded in S21 are completely filled into the corresponding weight and threshold storage locations in this temporary network computation graph. Next, the system reads the historical coal spontaneous combustion verification dataset from an independent data storage area. Each row of data in this dataset is called a sample, containing historical environmental parameters as input and the corresponding, actually observed and recorded coal spontaneous combustion temperature value as the target output. For each sample in the verification dataset, the system inputs the historical environmental parameters of the sample into the aforementioned temporarily configured network computation graph, initiates a complete forward propagation calculation process as in S5, and obtains the network's predicted temperature value for that sample.

[0074] S23: Fitness Value Calculation: After the temporary network completes a prediction calculation for all samples in the validation dataset, the system begins to calculate the prediction error. For each sample, the square of the difference between the network's predicted temperature value and the actual observed temperature value is calculated. Then, the average of this squared difference for all samples is calculated, which yields the mean squared error. The smaller this mean squared error value, the stronger the fitting ability of the temporary network, i.e., the candidate parameter set, to historical data, and the better the corresponding ant path. To transform the error into a larger and better evaluation metric, the system calculates the reciprocal of this mean squared error as the fitness value of the path constructed by the current ant.

[0075] S24: Optimal Path Identification: In one iteration, after all M ants have completed path construction, decoding, network configuration, and fitness calculation, the system scans all M fitness values ​​generated in this iteration. It finds the value with the largest value and records its corresponding ant path as the local optimal path for this round. Simultaneously, the system maintains the global historical optimal fitness value and its corresponding path. The system compares the local optimal fitness value found in this round with the global historical optimal fitness value. If the local optimal fitness value is higher, the previously recorded global optimal path is replaced with the local optimal path, and the global historical optimal fitness value is updated; otherwise, the original global optimal path record remains unchanged.

[0076] S25: Global Pheromone Evaporation: At the end of each iteration, before pheromone enhancement, a global evaporation operation is performed on the entire pheromone concentration matrix or list. Specifically, the pheromone value of each element in the pheromone matrix, i.e., each selectable path segment, is multiplied by a fixed coefficient less than 1 and greater than 0, such as 0.95. This operation proportionally reduces the pheromone concentration on all path segments, simulating the natural phenomenon of pheromone evaporation over time. Its purpose is to gradually diminish past poor search experiences and prevent the algorithm from getting trapped in local optima.

[0077] S26: Elite Path Pheromone Enhancement: After global evaporation, the system only allows specific elite ants to add pheromones to the path segments they have traversed. These elite ants typically include ants that constructed the locally optimal path in this round, and ants that constructed the historical globally optimal path. For each elite path, the system iterates through each path segment from its start point to its end point, representing each specific choice made by the ant in each parameter dimension. For each such path segment, the system calculates the pheromone increment based on the fitness value corresponding to the elite path. The higher the fitness value, the larger the calculated increment. This increment is then added to the pheromone value of the path segment after its current evaporation. This mechanism ensures that path segments corresponding to high-quality solutions receive stronger positive pheromone feedback, thereby attracting more ants to choose similar paths in subsequent iterations, achieving a concentrated search for regions of high-quality solutions.

[0078] Furthermore, in step S2, the step of calling the ant colony optimization program to optimize the initial connection weight matrix and initial neuron threshold vector of the pre-configured neural network model to generate an optimized neural network parameter set is specifically implemented through the following detailed steps:

[0079] S2a: Parameter Encoding and Search Space Discretization: All connection weights and neuron thresholds to be optimized in the neural network model are arranged in order to form a multidimensional optimization vector; a discrete and finite value search interval is defined for each parameter dimension in the multidimensional optimization vector, and the interval is evenly divided into N levels, where N is an integer greater than 1; the complete path of each ant is composed of the sequence of level numbers it selects in each parameter dimension, and the path uniquely corresponds to the specific candidate neural network parameter value;

[0080] S2b: Ant colony initialization and pheromone matrix construction: Initialize an ant colony of size M, where M is an integer greater than 1; initialize pheromone concentration values ​​for each of the N levels in each parameter dimension of the multidimensional optimization vector, thereby forming a multidimensional pheromone concentration matrix; at the same time, calculate a static heuristic pheromone value for each level, which is proportional to the probability that the parameter value represented by that level is within a preset reasonable prior range;

[0081] S2c: Iterative search main loop: Set the maximum number of iterations K, where K is an integer greater than 1, and enter the iteration loop;

[0082] S2d: Single Ant Path Construction: For each ant in the colony, select a level for each dimension in the order of the parameter dimensions. For the current dimension, the ant calculates the probability of selecting each level using a roulette wheel selection strategy based on the current pheromone concentration and heuristic information value of all N levels in that dimension. The level is randomly selected according to the calculated probability distribution and serves as the ant's path node in the current dimension. This process is repeated until all dimensions have been selected, thus forming a complete path for the ant.

[0083] S2e: Path evaluation and fitness calculation: For all paths constructed by M ants in the current iteration, decode them into candidate parameters and calculate their corresponding fitness values ​​according to the methods described in S21 to S23.

[0084] S2f: Pheromone Update: Following the methods described in S24 to S26, update the pheromone concentration matrix based on the path evaluation results of the current iteration;

[0085] S2g: Iteration Termination Judgment and Optimal Solution Output: When the number of iterations reaches K, or the fitness value of the global optimal path does not improve significantly in several consecutive iterations, the iteration loop is terminated; the global optimal path in the record is decoded into specific neural network connection weight values ​​and neuron thresholds, and this set of parameters is the optimized neural network parameter set.

[0086] In this embodiment, S2a: Parameter Encoding and Search Space Discretization: First, the system sorts and ranks the neural network parameters that need to be optimized. Assuming the neural network has W connection weights and B neuron thresholds to optimize, these (W+B) parameters are arranged in a predetermined order, such as layer order or neuron order, into a one-dimensional vector, called the optimization vector. The goal of the optimization search is to find the optimal real value for each element of this vector. To apply the ant colony algorithm, the continuous real number search problem needs to be discretized. For the i-th parameter in the optimization vector, the system sets a reasonable lower limit Li and upper limit Ui based on prior knowledge or experience. This interval [Li, Ui] is considered a meaningful range of values ​​for the parameter. Then, this continuous interval is evenly divided into N smaller intervals, where N is a preset integer, such as 10 or 20. Each smaller interval can be represented by its ordinal level j, where j ranges from 1 to N. Therefore, when optimizing the i-th parameter, the ants do not directly choose a real number, but rather select from the N levels. After all parameter dimensions have been discretized in this way, the entire search space constitutes a high-dimensional grid. The process of an ant constructing a complete path involves selecting a level j_i from N levels for each parameter dimension i. This sequence Path={j_1,j_2,...,j_(W+B)}, composed of level indices, corresponds to the complete candidate solution. When it is necessary to decode the path into specific parameter values, for the i-th parameter, the corresponding real value of the parameter can be calculated according to the selected level j_i. For example, the midpoint value between cells corresponding to this level can be used: Value_i=Li+(j_i-0.5)×(Ui-Li) / N.

[0087] S2b: Ant Colony Initialization and Pheromone Matrix Construction: Based on the preset ant colony size M, the system creates a data structure for M ant agents. Each ant agent contains its current path, fitness value, and other state information. Simultaneously, the system needs to initialize the pheromone environment. Since each parameter dimension i has N selectable levels, for (W+B) dimensions, a total of (W+B)×N selectable path segments need to be managed. The system initializes a two-dimensional matrix T or an equivalent data structure, with the number of rows equal to the total number of parameter dimensions (W+B) and the number of columns equal to the number of levels N. The initial values ​​of elements τ(i,j) in the matrix are set to the same positive constant, indicating that at the start of the search, the probability of an ant choosing the j-th level of parameter i is equal. Furthermore, the system can also calculate a heuristic information matrix H. The element η(i,j) in the matrix is ​​a static value that does not change with iteration. Its calculation can be based on prior knowledge; for example, if it is known empirically that a certain range of parameter i corresponding to a certain level j is more likely to lead to good network performance, a larger η(i,j) value can be set for that level to guide the ants in the early stages of the search. If there is no explicit prior knowledge, all η(i,j) can be set to 1.

[0088] S2c: Iterative Search Main Loop: The system sets a maximum iteration count counter K_max and a current iteration counter k, and sets k=1. An early termination condition can be set, such as if the global optimal solution has not improved for several consecutive iterations. Then, the main loop iterates through (while(k<=K_max and the early termination condition is not met)).

[0089] S2d: Inner Loop for Single Ant Path Construction: In each iteration k of the main loop, the system constructs a path for each ant in the colony, assuming its number is m. For ant m, path construction is a sequential decision-making process. It starts from the first parameter dimension (i=1) and ends at the last parameter dimension (i=W+B). For the parameter dimension i currently being decided, ant m needs to choose from N levels. It calculates the probability p(i,j) of choosing a level j based on the pheromone concentration τ(i,j) and heuristic information η(i,j) at each level j. The probability p(i,j) is proportional to the power of α of [τ(i,j)] multiplied by the power of β of [η(i,j)], where α and β are preset parameters that adjust the relative importance of pheromones and heuristic information. After calculating the probability values ​​for all N levels, the system normalizes them so that their sum is 1. Then, the algorithm generates a random number in the interval [0,1). The choice of level is determined by which interval within the cumulative probability intervals defined by the probabilities of each level falls into. Ant m records this choice. Then, the ant moves to the next parameter dimension (i+1) and repeats the probability-based selection process until all dimensions have been decided. This generates a complete path sequence for ant m.

[0090] S2e: Path evaluation and fitness calculation in parallel or sequential order: After all M ants have completed their path construction in this iteration, the system evaluates the path of each ant according to steps S21 to S23. This process can be performed sequentially for the M ants, or it can utilize the parallel capabilities of the processor to evaluate multiple ants simultaneously. The result of the evaluation is to calculate the fitness value fitness_m for each ant.

[0091] S2f: Pheromone Update: After the fitness of all ants has been calculated, the system performs optimal path identification, global pheromone evaporation, and elite path pheromone enhancement according to steps S24 to S26. Specifically, first, the local optimal path of this generation and the global optimal path to date are found; then, the pheromone matrix T is traversed, and an evaporation operation is performed on each element: τ(i,j)=ρ×τ(i,j), where ρ is an evaporation coefficient less than 1; finally, for the local optimal path and the global optimal path, the pheromone enhancement is performed on each decision point dimension i and the selected level j on the path: τ(i,j)=τ(i,j)+Δτ, where Δτ is Q_local×fitness_local for the local optimal path and Q_global×fitness_global for the global optimal path, where Q_local and Q_global are preset enhancement coefficients.

[0092] S2g: Iteration Termination Judgment and Optimal Solution Output: After completing the pheromone update, the current iteration k ends. The system increments the iteration counter k by 1 and checks whether the loop termination condition k>K_max or the global optimal fitness value has not been improved for multiple consecutive generations is met. If satisfied, the main loop exits; otherwise, it jumps to S2d to start the next generation iteration. When the main loop ends, the system outputs the recorded global optimal path. This path is converted into (W+B) specific real values ​​according to the decoding method described in S2a. This set of real values ​​is the optimized neural network parameter set obtained through the ant colony algorithm optimization search.

[0093] Furthermore, referring to Figure 3 In S3, S4, and S5, the construction, instantiation, and prediction calculation of the neural network model specifically include:

[0094] S3a: The neural network model is a feedforward neural network structure containing an input layer, at least a hidden layer, and an output layer; the number of neurons in the input layer is strictly equal to the number of parameters in the current environmental state parameter set, and each neuron receives the corresponding state parameter value; the output layer contains neurons used to output the predicted value of the spontaneous combustion temperature; the neurons between each layer are fully connected.

[0095] S3b: The values ​​in the optimized neural network parameter set generated in S2 are assigned to the corresponding connection weights and neuron thresholds in the neural network model according to the predefined mapping rules; whereby the connection weight values ​​are assigned to the weight coefficients of each directed edge connecting adjacent layer neurons, and the neuron thresholds are assigned to the internal bias terms of each neuron except the input layer.

[0096] S3c: When the current environmental state parameter set is input to the input layer, the data begins to propagate forward. For each neuron in the hidden and output layers, the calculation process is to calculate the weighted sum of the signal values ​​transmitted from all input connections of the neuron and the corresponding connection weights, add this weighted sum to the threshold of the neuron to obtain the net input value of the neuron, and map the net input value through a preset, non-linear activation function to obtain the output value of the neuron. The output value of the previous layer is used as the input signal of the next layer and is passed layer by layer.

[0097] S3d: The final output value of the output layer neuron, after being processed by a linear or preset scaling function, is determined as the predicted value of the spontaneous combustion temperature of the target coal sample at a preset time point in the future; the preset time point is one or more future time intervals that are pre-set according to the response time requirements of the early warning of coal spontaneous combustion.

[0098] In this embodiment, S3a: Neural Network Structure Definition: The neural network model adopts a feedforward, hierarchical computation structure. During the system design phase, its specific topology is determined: the number of neurons in the input layer, L_input, is determined based on the total number of parameters that must be included and may be expanded in the current environmental state parameter set. For example, if there are 5 basic parameters and 2 expanded scenario parameters, then L_input must be at least 7. Each input layer neuron is designed to receive and transmit only a specific type of parameter value. For example, the first neuron is dedicated to receiving the internal temperature of the coal body, the second neuron is dedicated to receiving the oxygen concentration, and so on. The number of hidden layers, L_hidden, and the number of neurons in each hidden layer, L_hidden_neuron, are pre-set according to the problem complexity. For example, it can be configured with two hidden layers, the first hidden layer containing 10 neurons and the second hidden layer containing 5 neurons. All neurons between adjacent layers implement a fully connected strategy, that is, each neuron in the l-th layer is connected to each neuron in the (l+1)-th layer through a directed connection. The number of neurons in the output layer is fixed at 1, corresponding to the spontaneous combustion temperature value to be predicted. Each neuron, except those in the input layer, has a pre-defined activation function. For example, neurons in the hidden layer use a sigmoid nonlinear function, while neurons in the output layer use a linear identity function. These structural parameters—the number of layers, the number of neurons per layer, the connection methods, and the type of activation function—remain fixed after the model is defined, constituting the architecture or topology of the neural network model.

[0099] S3b: Model Parameter Instantiation: The architecture of the neural network model only defines the data flow and nodes for computation; its specific computational behavior is determined by parameter weights and thresholds. In step S3, the system performs parameter instantiation. It reads the optimized neural network parameter set generated in step S2 from storage. This set is a long, ordered list of values. The system assigns these values ​​according to a predefined parameter-position mapping table that perfectly matches the model architecture. The mapping table explicitly specifies that: the 1st to W1st values ​​in the list correspond sequentially to the connection weights from the 1st neuron in the input layer to the 1st neuron in the first hidden layer, the 2nd neuron in the input layer to the 1st neuron in the first hidden layer, and so on, until all weights are filled. Following all the weight values, the subsequent values ​​in the list correspond to the thresholds of each neuron: first, the threshold of the first neuron in the first hidden layer; then the threshold of the second neuron in the first hidden layer, and so on, followed by the thresholds of each neuron in the second hidden layer, and finally the thresholds of the neurons in the output layer. The system strictly follows this mapping order, writing the values ​​in the list one by one into the memory addresses of the corresponding weight and threshold variables in the neural network model's computation graph. After this writing is complete, the neural network model transforms from an empty shell into a function instance capable of performing specific mathematical operations, with all computational parameters already determined.

[0100] S3c: Forward Propagation Calculation Process: When the instantiated model is invoked in step S5, it performs forward propagation calculation. The input data, the current environment state parameter vector, is fed into the input layer. The input layer neurons do not perform calculations; they directly pass the received values ​​as their output values ​​to the next layer, the first hidden layer. For each neuron in the first hidden layer, denoted as neuron_h, it performs the following operations: First, it initializes the accumulator sum to 0. Then, it iterates through all connections with the input layer neurons. For each connection from input layer neuron_i, it reads the weight value w_{i->h} from the connection weight storage area and simultaneously receives the output value x_i from input layer neuron_i. It calculates the product w_{i->h} × x_i and adds the result to sum. After iterating through all input connections, sum is the weighted sum. Next, neuron_h reads its threshold bias b_h from its own storage area. It calculates net_h = sum + b_h, which is the net input of this neuron. Finally, it inputs net_h into its built-in activation function F(), calculating its output value out_h = F(net_h). This out_h becomes the activation signal for this neuron, passed to the next hidden layer or output layer. Each neuron in the hidden layer and subsequent layers repeats the exact same process of collecting the outputs of all neurons in the previous layer, multiplying them by their corresponding connection weights, summing them, adding their own threshold, and finally mapping them through the activation function. The signal is thus passed and transformed layer by layer.

[0101] S3d: Prediction Generation: After layer-by-layer calculations through all hidden layers, the signal finally reaches the single neuron in the output layer. This neuron performs its own weighted summation and thresholding calculation to obtain the net input value net_out. Since the activation function of the output layer neuron is usually preset to a linear function or an identity function, its output out_out = G(net_out) is often net_out itself, or a simple linear scaling such as out_out = scale × net_out + offset, where scale and offset are preset constants used to adjust the network output value to the actual temperature dimensions and range. This out_out is the original prediction value calculated by the neural network. Finally, the system interprets this original prediction value as a temperature prediction for a preset future time point. The preset time point is set according to the early warning requirements when the model is deployed, for example, it can be set to 24 hours from the current time. The system can maintain a time mapping relationship internally to ensure that the output prediction value is correctly labeled or understood as the predicted temperature at that specific future time. For situations requiring predictions at multiple time points, multiple output neurons can be deployed, each corresponding to a different time offset. Their calculation process is parallel and independent of that of a single output neuron.

[0102] Furthermore, the method also includes an adaptive treatment step to address the differences in spontaneous combustion tendency of different coal types:

[0103] Before executing S1, obtain the coal type identification information of the target coal sample;

[0104] Based on the coal type identification information, a corresponding target parameter configuration group is selected from a plurality of pre-stored parameter configuration groups; wherein, different parameter configuration groups are associated with different coal types and at least include configuration parameters related to the ant colony algorithm optimization search in S2, the configuration parameters including the number of ants M, the number of iterations K, the pheromone volatile factor, and the search range of each parameter in the multidimensional optimization vector;

[0105] Use the parameters in the target parameter configuration group to initialize or adjust the running configuration of the ant colony algorithm optimization program, and then execute steps S2 to S6.

[0106] In this embodiment, before executing S1, the system first reads the coal type identification information of the target coal sample through a human-computer interaction interface or from an associated database. This identification information can be a standard coal type classification code, such as the code for anthracite, bituminous coal, or lignite, or it can be a coal mine or user-defined coal quality type label.

[0107] The system has a pre-stored coal type configuration database. This database uses coal type identification information as the primary key and stores multiple associated parameter configuration groups. Upon receiving the identification information of a target coal type, the system uses this as the primary key to query the database, retrieves a record that precisely matches the target coal type, and extracts the corresponding target parameter configuration group from that record. The target parameter configuration group is a data structure containing multiple key-value pairs, specifically designed to guide or adjust the subsequent optimization search process of the ant colony algorithm. These configuration parameters include, but are not limited to: the ant colony size M, the maximum number of iterations K, the pheromone volatility coefficient ρ, the importance factors α and β of pheromone and heuristic information, the pheromone enhancement coefficient Q, and most importantly, the search interval range [Li, Ui] for each weight and threshold parameter to be optimized in the neural network model. For different coal types, due to their significantly different spontaneous combustion characteristics, the optimal neural network parameters may be distributed across different numerical ranges; therefore, pre-setting differentiated search intervals for different coal types is crucial.

[0108] Before step S2 begins, the system calls the initialization module of the ant colony optimization program, passing the target parameter configuration group as input parameters to this module. The initialization module resets the algorithm's internal state variables and control parameters based on the parameters in the configuration group. Specifically, it re-executes a search space discretization process similar to S2a based on [Li, Ui] defined for each neural network parameter in the configuration group, establishing a discretized grid suitable for the current coal type. Simultaneously, operating parameters such as the number of ants and the number of iterations are also set accordingly. After completing these initialization tasks, the system then proceeds with steps S2 to S6 as described above. In this way, the entire optimization and prediction process is customized based on the specific knowledge of the current coal type, improving the targeting and accuracy of the prediction.

[0109] Furthermore, the method also includes specific application steps for underground goaf scenarios in coal mines:

[0110] The set of current environmental state parameters obtained in S1 additionally includes parameters of air leakage intensity in the goaf and parameters of the distribution characteristics of residual coal thickness in the goaf.

[0111] In S5, the hidden layer of the neural network model is configured to have at least two sub-hidden layers, wherein the first sub-hidden layer is specifically used to fuse the parameters of the internal temperature of the coal body, the ambient oxygen concentration, and the air leakage intensity of the goaf to calculate the comprehensive oxidation heating potential index; the second sub-hidden layer receives the comprehensive oxidation heating potential index and other environmental state parameters and performs the final feature integration.

[0112] The output predicted auto-ignition temperature is marked as corresponding to a specific goaf zone or coordinate location.

[0113] In this embodiment, the current environmental state parameter set obtained in S1 must include two additional parameters specific to the goaf area, in addition to the basic parameters: one is the average air leakage intensity of a specific area of ​​the goaf area estimated by wind speed sensors or differential pressure calculation models, which reflects the oxygen supply rate; the other is the average thickness or thickness distribution characteristic parameters of the residual coal seam in the area, such as maximum thickness and average thickness, obtained by borehole detection or ground-penetrating radar data inversion.

[0114] To effectively handle these unique parameters and capture their coupling relationship with the basic parameters, the hidden layer structure was specifically configured in the design phase of the neural network model, namely S3a. The hidden layer was explicitly designed to contain at least two functionally ordered sub-hidden layers. The first sub-hidden layer was given a specific functional focus: the input received by its neurons was limited to, or primarily weighted, the parameters of the coal body internal temperature, ambient oxygen concentration, and the air leakage intensity of the goaf. This sub-hidden layer deeply integrates these three parameters, which are strongly correlated with the coal-oxygen complex and heat accumulation process, through internal computational weighted summation and nonlinear transformation. Its output can be regarded as an abstract comprehensive oxidation heating potential index. The second sub-hidden layer acts as a general feature integrator. It receives the comprehensive oxidation heating potential index output from the first sub-hidden layer, along with other environmental state parameters, such as coal body porosity, ambient humidity, surrounding rock temperature, and residual coal thickness characteristics. The second sub-hidden layer performs secondary integration and transformation on all these input information, extracting the final high-level features used for temperature prediction.

[0115] When outputting the predicted spontaneous combustion temperature in S6, the predicted value will not exist in isolation because goaf early warning requires locating the potential hazard area. The system will bind this predicted value to the sensor deployment location or goaf zoning code on which the prediction was based. For example, the predicted value will be marked as corresponding to the 15-20 meter depth zone on the upper side of the No. 3 grouting roadway, or associated with the underground coordinates (X, Y, Z). This binding information can come from the mapping table of sensor node IDs and their physical locations in step S1, or it can be a user-specified assessment area identifier. The bound prediction information is more conducive to guiding targeted prevention and control measures.

[0116] Furthermore, the method also includes specific application steps for open-air coal storage yard scenarios:

[0117] The set of current environmental state parameters obtained in S1 additionally includes environmental wind speed and direction, solar radiation intensity, and coal pile surface temperature gradient parameters.

[0118] The method further includes a correction step after S5: based on the current seasonal information and historical meteorological statistics, a dynamic correction offset is applied to the initial spontaneous combustion temperature prediction value calculated by the neural network model. The correction offset is obtained by querying a predefined seasonal-meteorological correction table.

[0119] In this embodiment, the set of current environmental state parameters obtained in S1, for open-air coal piles, needs to include parameters reflecting the influence of external weather and solar radiation. Therefore, in addition to the basic parameters, it must also include: environmental wind speed and direction obtained from weather stations or sensors that can be synthesized into an effective wind speed perpendicular to the coal pile surface; solar radiation intensity measured by a total solar radiation meter; and surface temperature gradient calculated by an array of temperature sensors arranged at different depths on the surface of the coal pile, such as the temperature difference between the surface and 10 cm below the surface.

[0120] Considering that meteorological conditions, especially solar radiation and seasonal variations, directly affect the surface temperature of coal piles, and that this impact may not be fully covered by the static historical data of the neural network, this invention introduces a post-processing correction step. This step occurs after the S5 model calculation is completed and before the final output in S6. Specifically, the system first obtains the current date information to determine the season, such as spring, summer, autumn, or winter. Simultaneously, the system accesses a predefined seasonal-meteorological correction table. This table is a static data table; its row index may be the season or month, and its column index may be the weather type (sunny, cloudy, rainy) or solar radiation intensity range. Each cell stores an empirical temperature correction offset in degrees Celsius. This offset represents the typical temperature deviation on the coal pile surface caused by factors such as solar radiation and air convection under the given season and meteorological conditions, which may not have been fully learned by the neural network model. Based on the current season and the real-time solar radiation intensity, or in combination with wind speed, the system queries this correction table to obtain the specific dynamic correction offset ΔT_corr. Then, the initial auto-ignition temperature prediction value T_raw, directly calculated by the neural network in step S5, is added to this correction offset or combined using other calculation methods specified in the table to obtain the corrected final prediction value T_final = T_raw + ΔT_corr. This T_final is the value that will be output in step S6. This correction mechanism is equivalent to adding a compensation step based on physical experience and climatic laws to the prediction of open-air scenarios.

[0121] Furthermore, the method also includes specific application steps for coal transportation scenarios:

[0122] The set of current environmental state parameters obtained in S1 is collected in real time by a sensor network deployed on the transport vehicle, and additionally includes parameters such as the change rate of coal vibration compaction and the local gas residence time in the carriage.

[0123] In S5, the neural network model is periodically invoked in a rolling time window manner. Each time it is invoked, the latest set of environmental state parameters is input, and the predicted curve of spontaneous combustion temperature during the future transportation period is output.

[0124] The predicted auto-ignition temperature value is output together with the geographical location information or transportation mileage information.

[0125] In this embodiment, the current environmental state parameter set acquired in S1 is sourced from dynamically moving transport vehicles such as train carriages, cargo ship holds, and truck cargo boxes. A sensor network of temperature, gas, pressure, and vibration sensors is deployed on the vehicle, and the data is collected in real-time to a processor unit built into or accompanying the vehicle via wireless or wired means. In addition to the basic parameters, the collected parameters must include two additional parameters specific to the transportation process: first, the vibration frequency and amplitude of the coal during transportation, obtained through vibration sensor signal analysis, which estimates the rate of change in coal density and affects oxygen permeability; second, the average residence time of gas in the coal pile voids, estimated by calculating local airflow velocity and combining it with the carriage structure, which affects the accumulation of oxidation reaction gases.

[0126] Because the transportation environment is constantly changing, dynamic risk tracking is necessary. Therefore, in S5, the neural network model is periodically invoked and executed using a rolling time window. The system sets a fixed prediction period, for example, every 30 minutes. At the beginning of each period, the system collects the latest sensor data to form the latest set of current environmental state parameters. Then, it instantiates the neural network model, which may have been optimized using global historical data or optimized for the type of coal being transported. This latest parameter set is input, and a complete forward propagation calculation is performed. However, this calculation outputs not a single future prediction value, but a series of prediction values ​​for future time points, collectively forming a spontaneous combustion temperature prediction curve for the next transportation period, for example, the next 6 hours. This curve describes the predicted temperature trend under the assumption that future environmental conditions remain unchanged or follow the current trend.

[0127] In S6, the output auto-ignition temperature prediction value or curve needs to be correlated with the spatial location information of the transportation process to provide a dynamic risk geographic distribution. Therefore, the system binds the prediction results with real-time geographic location information provided by the transport vehicle, such as GPS coordinates or cumulative transport mileage information from the departure point to the present. For example, the output data packet can be formatted as: at mileage point KM255, the predicted maximum core temperature in the next 2 hours is T°C. This spatiotemporal binding output is very beneficial for the transportation dispatch center to grasp the risk status of coal carriages at different locations along the entire transportation route in real time, realizing dynamic monitoring and management.

[0128] Furthermore, the output of the predicted auto-ignition temperature value in step S6 specifically includes at least one of the following output formats:

[0129] Format 1: The predicted spontaneous combustion temperature is sent in digital form to the graphical user interface for real-time display, and a visual alarm is triggered when the predicted value exceeds a first threshold, and an auditory alarm is triggered when it exceeds a higher second threshold.

[0130] Format 2: The predicted spontaneous combustion temperature value, its corresponding target coal sample identifier, prediction time point, and data acquisition timestamp are encapsulated into a structured data message and sent to the upper-level monitoring system through an industrial communication interface;

[0131] Form 3: Compare the predicted value of the spontaneous combustion temperature with the historical predicted value and the actual observed value to generate a prediction reliability assessment index, and store the index and the predicted value together in a local log file.

[0132] In this embodiment, the local human-machine interface (GUI) display and alarm mechanism is as follows: The system is equipped with a graphical user interface (GUI). The processor formats the calculated predicted spontaneous combustion temperature value T_pred, along with its corresponding target name (e.g., coal pile No. 3) and the predicted time point, into a display string. Graphics library functions are called to refresh and display this information in real time in designated areas of the GUI, such as the main dashboard, data tables, and trend charts. Simultaneously, the system internally presets two temperature thresholds: a first-level warning threshold T_warn and a second-level alarm threshold T_alarm, where T_alarm > T_warn. The system continuously compares T_pred with these two thresholds. When T_pred first exceeds T_warn, the processor triggers a visual alarm, for example, by changing the displayed T_pred number in the GUI to flash yellow or by displaying a warning message box on the side of the interface. When T_pred further increases and exceeds the higher T_alarm, the processor, while maintaining the visual alarm, triggers an auditory alarm, i.e., by playing a pre-recorded warning sound or voice prompt through an audio device. This method is suitable for on-site duty room monitoring.

[0133] Method Two: Remote Data Reporting. The system is designed with industrial communication interfaces such as ModbusTCP, OPCUA, and HTTPAPI. The processor encapsulates the predicted auto-ignition temperature value T_pred, along with necessary metadata including the unique identifier ID of the target coal sample, the future time point targeted for the prediction (timestamp_future), and the data collection timestamp (timestamp_sample) used for this prediction calculation, according to a predefined message format, generating a structured data message such as a JSON object or binary data packet. Then, through the industrial communication interface, according to the agreed communication protocol, the message is actively sent to the designated receiving endpoint of the superior monitoring system, cloud platform, or data center. This method is suitable for integration into large-scale centralized monitoring networks.

[0134] Form 3: Local Log Recording and Self-Evaluation: The system allocates a dedicated log file in memory. After each prediction execution, the processor not only records the current predicted value T_pred, but also retrieves several previous predicted values ​​for the same monitoring point from the historical record, as well as the actual observed temperature values, if applicable. By comparing the continuous predicted values ​​of the sequence, its volatility can be calculated; by comparing historical predicted values ​​with actual observed values, a simple consistency index, such as the reciprocal of the mean absolute error of the most recent N predictions, can be calculated. Based on these calculations, the system generates a prediction reliability evaluation index between 0 and 1. Finally, the processor appends the current timestamp, T_pred, and this reliability index as a complete record entry to the local log file. This form provides process tracking and quality self-checking capabilities, facilitating subsequent analysis and auditing.

[0135] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0136] Furthermore, the functional units in the various embodiments of this application can be integrated into the processing unit, or each unit can exist physically separately, or two or more units can be integrated into the unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

[0137] The specific embodiments of the invention have been described in detail above, but they are only examples, and this application is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications or substitutions to the invention are also within the scope of this application. Therefore, all equivalent changes, modifications, and improvements made without departing from the spirit and principles of this application should be covered within the scope of this application.

Claims

1. A method for predicting the spontaneous combustion temperature of coal based on an ant colony algorithm-optimized neural network, characterized in that, include: S1: Obtain the current environmental state parameter set of the target coal sample, which includes the internal temperature of the coal body, the ambient oxygen concentration, the coal body porosity, the ambient humidity, and the temperature of the surrounding rock where the coal body is located. S2: Call the ant colony algorithm optimization program to optimize and search the initial connection weight matrix and initial neuron threshold vector of the pre-configured neural network model to generate an optimized neural network parameter set; S3: Load the optimized neural network parameter set into the corresponding structure of the neural network model to complete the instantiation of the neural network model; S4: Input the current environment state parameter set obtained in S1 into the instantiated neural network model; S5: Run the instantiated neural network model to perform calculations. The hidden layer of the neural network model performs nonlinear transformation and feature fusion on the input parameters. The output layer calculates the predicted value of the spontaneous combustion temperature of the target coal sample at a future preset time point based on the output result of the hidden layer and outputs the predicted value of the spontaneous combustion temperature. In S2, the optimization search includes: constructing a search space that represents candidate neural network parameter solutions using individual ant paths; calculating the state transition probability of ants based on pheromone concentration and heuristic information on the path to explore new solutions; evaluating the quality of ant paths based on the fitting accuracy of the historical coal spontaneous combustion dataset after the candidate neural network parameter solutions are input into the neural network model; dynamically updating the pheromone concentration on the path based on the evaluation result; guiding the ant colony to gather in a region of better parameter solutions; and after multiple iterations, determining the parameter solution corresponding to the globally optimal path as the optimized neural network parameter set. The process of evaluating the fitting accuracy of the ant path to the historical coal spontaneous combustion dataset based on the parameter solutions of the candidate neural network inputted into the neural network model, and dynamically updating the pheromone concentration on the path based on the evaluation results, specifically includes: S21: For each ant's complete path constructed in the current iteration, decode the path into complete candidate neural network parameters, including the connection weights between all layers and the thresholds of each neuron; S22: Assign the candidate neural network parameters to a feedforward neural network with a fixed structure, and perform forward propagation calculation on the network using an independent historical coal spontaneous combustion verification dataset. The historical coal spontaneous combustion verification dataset contains multiple sets of historical environmental state parameters and their corresponding, subsequently observed coal spontaneous combustion temperatures. S23: Calculate the mean square error between the predicted temperature and the actual observed temperature of all samples in the historical coal spontaneous combustion verification dataset by the network, and use the reciprocal of the mean square error as the fitness value for evaluating the quality of the current ant path. S24: After all ants have completed path construction in one iteration, find the local optimum with the highest fitness value in this iteration and the global optimum with the highest fitness value in all iterations so far; S25: Perform a global evaporation operation on all paths of pheromones to simulate the natural evaporation of pheromones; S26: Only ants corresponding to the local optimal path and the global optimal path are allowed to add pheromones on the path segments they pass through, and the intensity of the added pheromones is proportional to the fitness value of the corresponding path. The method also includes specific application steps for underground goaf scenarios in coal mines: The set of current environmental state parameters obtained in S1 additionally includes parameters of air leakage intensity in the goaf and parameters of the distribution characteristics of residual coal thickness in the goaf. In S5, the hidden layer of the neural network model is configured to have at least two sub-hidden layers, wherein the first sub-hidden layer is specifically used to fuse the parameters of the internal temperature of the coal body, the ambient oxygen concentration, and the air leakage intensity of the goaf to calculate the comprehensive oxidation heating potential index; the second sub-hidden layer receives the comprehensive oxidation heating potential index and other environmental state parameters and performs the final feature integration. The output predicted auto-ignition temperature is marked as corresponding to a specific goaf zone or coordinate location.

2. The method according to claim 1, characterized in that, In step S2, the ant colony optimization program is invoked to optimize the initial connection weight matrix and initial neuron threshold vector of the pre-configured neural network model to generate an optimized neural network parameter set. This is achieved through the following detailed steps: S2a: Parameter Encoding and Search Space Discretization: All connection weights and neuron thresholds to be optimized in the neural network model are arranged in order to form a multidimensional optimization vector; a discrete and finite value search interval is defined for each parameter dimension in the multidimensional optimization vector, and the interval is evenly divided into N levels, where N is an integer greater than 1; the complete path of each ant is composed of the sequence of level numbers it selects in each parameter dimension, and the path uniquely corresponds to the specific candidate neural network parameter value; S2b: Ant colony initialization and pheromone matrix construction: Initialize an ant colony of size M, where M is an integer greater than 1; initialize pheromone concentration values ​​for each of the N levels in each parameter dimension of the multidimensional optimization vector, thereby forming a multidimensional pheromone concentration matrix; at the same time, calculate a static heuristic pheromone value for each level, which is proportional to the probability that the parameter value represented by that level is within a preset reasonable prior range; S2c: Iterative search main loop: Set the maximum number of iterations K, where K is an integer greater than 1, and enter the iteration loop; S2d: Single Ant Path Construction: For each ant in the colony, select a level for each dimension in the order of the parameter dimensions. For the current dimension, the ant calculates the probability of selecting each level using a roulette wheel selection strategy based on the current pheromone concentration and heuristic information value of all N levels in that dimension. The level is randomly selected according to the calculated probability distribution and serves as the ant's path node in the current dimension. This process is repeated until all dimensions have been selected, thus forming a complete path for the ant. S2e: Path evaluation and fitness calculation: For all paths constructed by M ants in the current iteration, decode them into candidate parameters and calculate their corresponding fitness values ​​according to the method described in S21 to S23 of claim 1. S2f: Pheromone Update: According to the method described in S24 to S26 of claim 1, update the pheromone concentration matrix based on the path evaluation results of the current iteration; S2g: Iteration Termination Judgment and Optimal Solution Output: When the number of iterations reaches K, or the fitness value of the global optimal path does not improve significantly in several consecutive iterations, the iteration loop is terminated; the global optimal path in the record is decoded into specific neural network connection weight values ​​and neuron thresholds, and this set of parameters is the optimized neural network parameter set.

3. The method according to claim 1 or 2, characterized in that, In S3, S4, and S5, the construction, instantiation, and prediction calculation of the neural network model specifically include: S3a: Neural Network Structure Definition: The neural network model is a feedforward neural network structure containing an input layer, at least a hidden layer, and an output layer; the number of neurons in the input layer is strictly equal to the number of parameters in the current environmental state parameter set, and each neuron receives the corresponding state parameter value; the output layer contains neurons for outputting the predicted value of the spontaneous combustion temperature; the neurons between each layer are fully connected. S3b: Model parameter instantiation: The values ​​in the optimized neural network parameter set generated in S2 are assigned to the corresponding connection weights and neuron thresholds in the neural network model according to the predefined mapping rules; wherein, the connection weight values ​​are assigned to the weight coefficients of each directed edge connecting adjacent layer neurons, and the neuron thresholds are assigned to the internal bias terms of each neuron except the input layer. S3c: Forward Propagation Calculation Process: When the current environment state parameter set is input to the input layer, the data begins forward propagation; for each neuron in the hidden layer and output layer, the calculation process is to calculate the weighted sum of the signal values ​​transmitted from all input connections of the neuron and the corresponding connection weights, add this weighted sum to the threshold of the neuron to obtain the net input value of the neuron, and map the net input value through a preset, non-linear activation function to obtain the output value of the neuron; the output value of the previous layer is used as the input signal of the next layer, and is passed layer by layer; S3d: Prediction value generation: The final output value of the output layer neuron, after being processed by a linear or preset scaling function, is determined as the predicted value of the spontaneous combustion temperature of the target coal sample at a preset time point in the future; the preset time point is one or more future time intervals pre-set according to the response time requirements of early warning of coal spontaneous combustion.

4. The method according to claim 1, characterized in that, The method also includes adaptive processing steps to address the differences in spontaneous combustion tendency of different coal types: Before executing S1, obtain the coal type identification information of the target coal sample; Based on the coal type identification information, a corresponding target parameter configuration group is selected from a plurality of pre-stored parameter configuration groups; wherein, different parameter configuration groups are associated with different coal types and at least include configuration parameters related to the ant colony algorithm optimization search in S2, the configuration parameters including the number of ants M, the number of iterations K, the pheromone volatile factor, and the search range of each parameter in the multidimensional optimization vector; Use the parameters in the target parameter configuration group to initialize or adjust the running configuration of the ant colony algorithm optimization program, and then execute steps S2 to S5.

5. The method according to claim 1, characterized in that, The predicted auto-ignition temperature value output in step S5 specifically includes at least one of the following output formats: Format 1: The predicted spontaneous combustion temperature is sent in digital form to the graphical user interface for real-time display, and a visual alarm is triggered when the predicted value exceeds a first threshold, and an auditory alarm is triggered when it exceeds a higher second threshold. Format 2: The predicted spontaneous combustion temperature value, its corresponding target coal sample identifier, prediction time point, and data acquisition timestamp are encapsulated into a structured data message and sent to the upper-level monitoring system through an industrial communication interface; Form 3: Compare the predicted value of the spontaneous combustion temperature with the historical predicted value and the actual observed value to generate a prediction reliability assessment index, and store the index and the predicted value together in a local log file.

Images