Intelligent downhole simulation fire training system

Abstract

The application discloses an intelligent underground fire simulation training system, and relates to the technical field of fire simulation training.The system comprises a data acquisition module, a man-machine interaction module, a disaster simulation module and an intelligent processing control module.The mine structure attribute information is mapped into a feature vector by an encoder, and a first machine learning model is injected in a gated fusion manner, so that the generated training scheme is always consistent with the actual risk image of the target mine.When the system is deployed to a new mine, only the mine structure feature vector needs to be updated to realize rapid adaptation without retraining the entire model.Based on the trained first machine learning model, the optimal training scheme that can maximize the improvement of the comprehensive ability of the trainee is searched in the candidate training scheme sequence space, so that the training effect of the trainee in the underground fire training is improved according to the learning condition and the individualized customized training scheme of the trainee.

Description

Technical Field

[0001] This invention relates to the field of fire simulation training technology, and more specifically to an intelligent underground fire simulation training system. Background Technology

[0002] Mine fires are a major threat to coal mine safety, characterized by their suddenness, rapid spread, and difficulty in rescue. To improve the fire emergency response capabilities of underground workers, mine fire simulation training systems are widely used in safety training.

[0003] Existing underground fire simulation training systems often employ simple emergency shutdown or manual reset methods for their safety reset functions, resulting in low efficiency for continuous training. They typically follow a pre-set, fixed training curriculum, failing to dynamically adjust training plans based on trainees' historical performance, skill development patterns, and the actual risk characteristics of the target mine. This leads to problems such as low training relevance and efficiency. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent downhole simulated fire training system to address the aforementioned shortcomings of the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent downhole simulated fire training system, comprising a data acquisition module, a human-computer interaction module, a disaster simulation module, and an intelligent processing and control module;

[0006] The data acquisition module is used to synchronously collect trainee behavior data and environmental status data during the training process;

[0007] The human-computer interaction module is used to visualize data and obtain user feedback information;

[0008] The disaster simulation module has multiple simulations covering various fire types;

[0009] The intelligent processing and control module is used to perform the following steps:

[0010] S1. By using a pre-constructed knowledge graph that integrates expert rules and historical cases, the fire risk level of each structural feature of the mine is inferred based on the structural attribute information of the target mine, and an initial training scheme is generated based on the fire risk level.

[0011] S2. During simulated fire training, the student behavior data and environmental status data collected in real time by the data acquisition module are input into the pre-trained scoring mapping model to generate real-time student scores.

[0012] S3. Input the historical training time series data of each trainee into the pre-trained first machine learning model, and use the mine structure attribute information input into the encoder obtained by the pre-trained first machine learning model as the conditional input of the first machine learning model to predict the learning ability growth curve of the trainee under different fire types and training frequencies; furthermore, similarly, the learning ability growth curve of the trainee under different regions-fire types and training frequencies can also be predicted.

[0013] S4. Initialize multiple candidate training scheme sequences and input them together with the historical training time series data and mine structure feature vector of the corresponding trainees into the first machine learning model to obtain the predicted score of the trainee under each candidate sequence; calculate the objective function value based on the predicted score, with the goal of maximizing the improvement of comprehensive ability, and determine the optimal candidate training scheme sequence through iterative search to obtain the trainee's simulated training plan sequence;

[0014] S5. Based on the student simulation training plan sequence, generate control signals and send them to the corresponding disaster simulation module to control the disaster simulation module to automatically execute fire simulation.

[0015] Furthermore, the intelligent processing control module also includes the following steps:

[0016] Before each simulated fire training exercise, a global safety initialization is performed. After power-on, a global reset command with the highest priority is sent to all disaster simulation modules. After collecting reset completion confirmation signals returned by all disaster simulation modules, the system enters standby mode.

[0017] Furthermore, the intelligent processing control module also includes the following steps:

[0018] During the simulated fire training, the trainees' real-time scores are input into the first machine learning model for optimization calculation, and the subsequent parts of the trainees' simulated training plan sequence are adjusted based on the results of the optimization calculation.

[0019] Furthermore, S1 includes the following steps:

[0020] The system receives structural and static attribute information of the target mine from the user's input through the human-computer interaction module, and obtains the structural attribute information of the target mine. The structural information includes, but is not limited to, roadway layout, ventilation network topology, and working face distribution. The static attribute information includes, but is not limited to, gas level, coal dust explosion index, and main combustible material types.

[0021] Based on the structural attribute information of the target mine, in a knowledge graph consisting of the knowledge of the expert group and historical fire cases of mines similar to the target mine, with the probability and severity of the occurrence of corresponding fire types, the probability and severity of the occurrence of fire types in each region that match the structural attribute information of the target mine are retrieved by a rule engine or graph query language. The risk level of each fire type in each region of the target mine is calculated by weighting according to preset probability weights and severity weights, and the region-fire type-risk level sequence of the target mine is generated in descending order of risk level.

[0022] The initial fire scenario training sequence is determined based on the order of the region-fire type-risk level sequence, and an initial training scheme is generated.

[0023] Furthermore, the pre-trained rating mapping model in S2 is pre-trained through the following steps:

[0024] Using the student's historical behavior data sequence and corresponding environmental state data sequence as input, and the student's historical expert evaluation score as output, a deep neural network model is trained to obtain a score mapping model.

[0025] Furthermore, the first machine learning model and encoder in S3 are jointly trained through the following steps:

[0026] For each trainee, the mine area, fire type, training time interval, and trainee real-time score corresponding to each simulated fire training session are collected in chronological order to generate the trainee's historical training time sequence data. The training time interval is calculated based on the training time of two adjacent simulated fire training sessions, that is, the difference between the training time of the current training record and the training time of the immediately preceding training record. The training time interval of the first training session is set to a preset initial value.

[0027] Historical training time series data is input into the sequence coding network of the first machine learning model to obtain a sequence coding vector; the sequence coding network is a neural network module used to extract temporal dependencies and contextual representations from a variable-length temporal input, and is used to compress the temporal patterns in the trainee's historical training time series data into a fixed-dimensional vector representation;

[0028] The target mine structure attribute information is input into the encoder to obtain the target mine structure attribute feature vector; the encoder is jointly trained with the first machine learning model;

[0029] The sequence encoding vector and the mine structural attribute feature vector are input into the gated fusion layer, concatenated, and then a gated vector is generated through a fully connected network. The gated vector is used to perform element-wise weighted combination of the sequence encoding vector and the mine structural attribute feature vector to obtain the fused feature vector.

[0030] The fused feature vector is input into the prediction head network of the first machine learning model, and the prediction head network outputs the prediction score of each fire type in the trainee’s future fire simulation training for a preset number of times. The prediction head network is a fully connected network module that connects the fused features and the prediction target. The output dimension is determined by the number of fire types and the prediction step size.

[0031] Using the real-time scores of the corresponding trainees in subsequent fire simulation training as the true labels, a weighted loss function between the predicted scores and the true labels is calculated, and the parameters of the sequence coding network, the gated fusion layer, the prediction head network, and the encoder of the first machine learning model are updated simultaneously through backpropagation.

[0032] Furthermore, the learning ability growth curve in S3 is composed of the predicted scores of each fire type output by the first machine learning model, arranged in the order of training for each student in a preset number of training sessions; wherein, each predicted score represents the expected performance level of the student for the corresponding fire type under the corresponding training order.

[0033] Furthermore, S4 includes the following steps:

[0034] Initialize multiple candidate training scheme sequences. Each candidate training scheme sequence consists of scheme items arranged in chronological order for a preset number of future training sessions. Each scheme item includes at least the mine area identifier, fire type, and training time interval.

[0035] For each candidate training scheme sequence, the student's historical training time series data and the candidate training scheme sequence are concatenated in the time dimension to form a joint input sequence. Each time step of the historical training time series data includes the student's real-time score, and the position corresponding to the student's real-time score in each time step of the candidate training scheme sequence is a placeholder value. Each time step of the concatenated historical training time series data and the candidate training scheme sequence also includes a type mask used to distinguish between the historical training time series data and the candidate training scheme sequence.

[0036] The joint input sequence and the mine structure feature vector are input into the first machine learning model, and the first machine learning model outputs the trainee's prediction score for each fire type at each time step corresponding to the candidate training scheme sequence.

[0037] Based on the predicted score, calculate the objective function value that characterizes the improvement of the student's comprehensive ability under the candidate training scheme sequence;

[0038] With the objective function value as the optimization objective, the candidate training scheme sequence is updated through iterative search. The candidate training scheme sequence with the optimal objective function value under preset constraints is used as the student simulation training customized plan.

[0039] Compared with existing technologies, the present invention provides an intelligent underground fire simulation training system that maps mine structural attribute information into feature vectors through an encoder and injects them into the first machine learning model in a gated fusion manner. This ensures that the generated training scheme is always consistent with the actual risk profile of the target mine. When the system is deployed to a new mine, it can be quickly adapted by simply updating the mine structural feature vectors, without having to retrain the entire model.

[0040] By searching the candidate training scheme sequence space for the optimal training scheme that maximizes the improvement of trainees' comprehensive abilities based on the first trained machine learning model, the training effect of underground fire training for trainees is improved by customizing training schemes according to trainees' learning situation and personalization.

[0041] By forcibly performing a global safety initialization after each system power-on, sending the highest-priority global reset command to all modules, and providing alarm prompts for unready modules, maintenance personnel can accurately locate faulty modules and perform targeted repairs. This improves the maintainability and operational efficiency of the system, effectively eliminating the safety risks introduced by abnormal termination of the last training or uncertain status after equipment maintenance. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0043] Figure 1 This is a system structure block diagram provided for an embodiment of the present invention. Detailed Implementation

[0044] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0045] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0046] Exemplary embodiments will be described more fully below with reference to the accompanying drawings; however, these exemplary embodiments may be embodied in different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will enable those skilled in the art to fully understand the scope of this disclosure.

[0047] Where there is no conflict, the various embodiments of this disclosure and the features thereof in the embodiments may be combined with each other.

[0048] As used herein, the term “and / or” includes any and all combinations of one or more related enumerated entries.

[0049] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. As used herein, the singular forms “a” and “the” are also intended to include the plural forms unless the context clearly indicates otherwise. It will also be understood that when the terms “comprising” and / or “made of” are used in this specification, the presence of the stated feature, integral, step, operation, element, and / or component is specified, but the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof is not excluded.

[0050] The embodiments described herein can be described with reference to plan views and / or cross-sectional views using the ideal schematic diagrams of this disclosure. Therefore, the example illustrations can be modified according to manufacturing techniques and / or tolerances. Therefore, the embodiments are not limited to those shown in the drawings, but include modifications to configurations formed based on manufacturing processes. Therefore, the areas illustrated in the drawings are schematic in nature, and the shapes of the areas shown in the figures illustrate specific shapes of areas of an element, but are not intended to be limiting.

[0051] Unless otherwise specified, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art. It will also be understood that terms such as those defined in commonly used dictionaries should be interpreted as having the same meaning as they have in the context of the relevant art and this disclosure.

[0052] Please see Figure 1 An intelligent downhole fire simulation training system includes a data acquisition module, a human-computer interaction module, a disaster simulation module, and an intelligent processing and control module.

[0053] The data acquisition module is used to synchronously collect trainee behavior data and environmental status data during training. The data acquisition module includes multiple sensors, each of which uploads data to the intelligent processing and control module via a wireless gateway at a preset frequency (e.g., 50Hz).

[0054] The student behavior data includes location information collected through location tags worn by students;

[0055] Physiological indicators (such as heart rate, body temperature, and blood oxygen saturation) collected by physiological sensors; and operational action data (such as simulating the lifting of a fire extinguisher, opening and closing of valves, and the use of rescue equipment) collected by operational sensors deployed in the training area.

[0056] Environmental status data includes environmental parameters collected by temperature sensors, smoke concentration sensors, and gas concentration sensors deployed in various training areas.

[0057] The human-computer interaction module is used to visualize data and obtain user feedback. Specifically, the human-computer interaction module may include a display screen, a touch panel, a voice broadcast device, and operation buttons.

[0058] The disaster simulation module includes multiple simulations covering various fire types. These modules include, but are not limited to: roof fall simulation module, rockfall simulation module, fire source simulation module, smoke simulation module, thermal radiation simulation module, sonic and optical explosion simulation module, wall condensation simulation module, and ceiling dripping water simulation module. Each disaster simulation module is equipped with a local controller and actuators, communicating with the intelligent processing and control module via a CAN bus or industrial Ethernet. It receives control commands and returns status confirmation signals containing sensor detection information such as limit switch status, encoder position, or valve body position. The actuators in each disaster simulation module utilize conventional technologies in the field. For example, the roof fall simulation module uses a motor-driven lead screw to simulate the translation and repositioning of rock blocks, and the smoke simulation module uses a solenoid valve to control the release of smoke generators. Their internal structures will not be described in detail here.

[0059] The intelligent processing control module is used to perform the following steps:

[0060] S1. Using a pre-constructed knowledge graph that integrates expert rules and historical cases, the fire risk level of each structural feature of the target mine is inferred based on the structural attribute information of the mine. An initial training scheme is generated based on the fire risk level, including the following steps:

[0061] S1.1 Receive the structural and static attribute information of the target area mine input by the user through the human-computer interaction module, and obtain the structural attribute information of the target mine; wherein, the structural information includes but is not limited to roadway layout, ventilation network topology, working face distribution, etc.; the static attribute information includes but is not limited to gas level, coal dust explosion index, main combustible material types, etc.

[0062] S1.2 Based on the structural attribute information of the target mine, in a knowledge graph consisting of the knowledge of the expert group and historical fire cases of mines similar to the target mine, with the probability and severity of the occurrence of corresponding fire types, the probability and severity of the occurrence of fire types in each region that match the structural attribute information of the target mine are retrieved by a rule engine (such as a production rule engine based on the Rete algorithm) or a graph query language (such as Cypher, SPARQL). Combining the rule reasoning and the statistical results of historical instances, the risk level of each fire type in each region of the target mine is calculated according to the preset probability weight and severity weight, and the region-fire type-risk level sequence of the target mine is generated in descending order of risk level.

[0063] The knowledge graph is constructed as follows: the expert group summarizes the fire risk experience of different areas of the mine into several "area-fire type" association rules. Each rule is accompanied by the probability weight and severity level of the consequences derived from the expert experience and historical accident statistics. At the same time, the expert group collects fire accident cases of similar mines over the years, extracts the mine structure parameters, fire area, fire type and consequences of each accident as instance nodes in the knowledge graph, and establishes associations with the corresponding area and fire type nodes.

[0064] S1.3 Determine the initial fire scenario training sequence based on the order of region-fire type-risk level sequence, and generate the initial training scheme.

[0065] In one embodiment, a global safety initialization is performed before each simulated fire training session. Upon power-up, a high-priority global reset command is sent to all disaster simulation modules, and a standby state is entered after collecting reset completion confirmation signals from all modules. If reset completion confirmation signals from all modules are not collected within a preset time, an alarm is triggered via the human-machine interface module, displaying information about unready modules.

[0066] S2. During simulated fire training, the student behavior data and environmental status data collected in real time by the data acquisition module are input into the pre-trained scoring mapping model to generate real-time scores for the students.

[0067] The pre-trained rating mapping model is pre-trained through the following steps:

[0068] Using historical student behavior data sequences and corresponding environmental state data sequences as input, and expert evaluation scores corresponding to the students' past performance as output, a deep neural network model is trained to obtain a score mapping model. Specifically, the student behavior data sequences and corresponding environmental state data sequences are aligned and concatenated with a unified timestamp before input, enabling the model to learn the causal relationship between actions and environmental changes. After necessary data preprocessing (such as denoising and normalization), the data is input into the pre-trained score mapping model. The Adam optimizer can be used during training, with mean squared error as the loss function.

[0069] S3. Input the historical training time-series data of each trainee into the pre-trained first machine learning model. Use the mine structure attribute information as the input to the encoder obtained by jointly training the first machine learning model with the pre-trained first machine learning model, and use this as the conditional input to predict the trainee's learning ability growth curve under different fire types and training frequencies. The learning ability growth curve is composed of the predicted scores of the trainee for each fire type in a preset number of future training sessions, arranged in training order, output by the first machine learning model. Each predicted score represents the trainee's expected performance level for the corresponding fire type under the corresponding training order.

[0070] The first machine learning model and encoder are jointly trained through the following steps:

[0071] (1) For each trainee, the mine area, fire type, training time interval and trainee real-time score corresponding to each simulated fire training are collected in chronological order to generate the trainee's historical training time series data; the training time interval is calculated based on the training time of two adjacent simulated fire trainings, that is, the difference between the training time of the current training record and the training time of the previous training record; the training time interval of the first training is set to a preset initial value; the mine area and fire type of each training record are converted into dense embedding vectors through the embedding layer, and the training time interval and trainee real-time score are normalized and used as continuous features. The features are spliced ​​together to form the input feature vector of the time step.

[0072] (2) Input the historical training time series data into the sequence encoding network of the first machine learning model to obtain the sequence encoding vector. The sequence encoding network is a neural network module used to extract temporal dependencies and contextual representations from variable-length temporal inputs, and is used to compress the temporal patterns in the trainee's historical training time series data into a fixed-dimensional vector representation. The sequence encoding network can adopt any neural network architecture that can process sequence data, such as recurrent neural network (RNN, including LSTM or GRU), Transformer encoder, temporal convolutional network (TCN). Taking the Transformer encoder as an example: it is composed of multiple multi-head self-attention sub-layers and feedforward fully connected sub-layers stacked together, with each sub-layer followed by residual connections and layer normalization. For example, the variable-length sequence can be compressed into a fixed-length sequence encoding vector by taking the mean pooling method on the sequence dimension of the Transformer encoder output.

[0073] (3) Input the target mine structure attribute information into the encoder to obtain the target mine structure attribute feature vector; the encoder is jointly trained with the first machine learning model; wherein, the encoder fuses and encodes the information of the roadway topology (processed by graph neural network or convolutional neural network) and static attributes (processed by fully connected network), and finally outputs a feature vector with fixed dimensions (such as 128 dimensions).

[0074] (4) The sequence encoding vector and the mine structure attribute feature vector are input into the gated fusion layer. After concatenation, a gate vector is generated through a fully connected network. The gate vector is then used to perform element-wise weighted combination of the sequence encoding vector and the mine structure attribute feature vector to obtain the fused feature vector. The specific operation is as follows: First, the sequence encoding vector and the mine structure attribute feature vector are concatenated. A fully connected network is used to map the concatenated vector to a gate vector with the same dimension as the original vector. Each element of the gate vector is activated by the sigmoid function and takes a value between 0 and 1. Then, the gate vector is used to perform element-wise weighted combination of the sequence encoding vector and the mine structure attribute feature vector. That is, each element of the gate vector controls the mixing ratio of the fused feature vector taken from the sequence encoding vector and the mine structure attribute feature vector in the corresponding dimension. The gate vector is used to adaptively adjust the contribution of the trainee's historical behavior encoding and the mine structure prior information in the fused representation at the element level. When trainees' historical training data is sparse, the model learns a smaller gate value, causing the fused features to rely more on mine structure feature vectors for prediction. When trainees' historical data is abundant and stable personal ability patterns are formed, the model learns a larger gate value, causing the fused features to retain the trainees' own behavioral encoding. This achieves a smooth transition from relying on group priors to highly personalized prediction.

[0075] (5) Input the fused feature vector into the prediction head network of the first machine learning model and output the prediction score of each fire type in the future preset number of fire simulation training sessions for the trainee; the prediction head network is a fully connected network module that connects the fused features and the prediction target, and the output dimension is determined by the number of fire types and the prediction step size.

[0076] (6) Using the real-time scores of trainees in subsequent fire simulation training as the true labels, calculate the weighted loss function between the predicted scores and the true labels, and simultaneously update the parameters of the sequence encoding network, gated fusion layer, prediction head network, and encoder parameters of the first machine learning model through backpropagation. The weighted loss function adopts the weighted mean square error. Since recent predictions have greater reference value for the generation of subsequent training plans, the time step weights can adopt an exponential decay form, so that the model pays more attention to the accuracy of recent predictions during training. The fire type weights are derived from the assessment results of the occurrence frequency and severity of each fire type in the target mine. Fire types with higher occurrence probabilities and more severe consequences are given greater weights, guiding the model to prioritize ensuring the accuracy of predictions for high-risk fire types. Backpropagation can use the AdamW optimizer, Adam optimizer, or RAdam optimizer.

[0077] During training, the input sequence can be the historical records of each student's most recent N training sessions. The predicted label is the student's actual scores for each fire type in the M subsequent training sessions following the most recent one (i.e., the student's real-time scores). For students with fewer than N training sessions, the sequence is padded with a zero vector, and a padding mask vector is appended to mark the boundary between valid data and the padding, ensuring the model ignores the padding during computation. To improve the model's generalization ability, the following data augmentation strategies can be applied to the input sequence during training: randomly discarding single training records to simulate data loss scenarios; applying small Gaussian noise perturbations to the scores to improve the model's robustness to score fluctuations; and randomly swapping the order of two consecutive training sessions of the same type, provided the time interval does not exceed a preset threshold, to enhance the model's adaptability to slight changes in the training order.

[0078] During system operation, when the number of newly added historical training time series data reaches 5%-15% or more of the current total training set, an incremental model fine-tuning is triggered. After fine-tuning, the model is evaluated on the validation set. If the performance degradation exceeds a preset threshold, the model is automatically rolled back to the weights before fine-tuning and an anomaly log is recorded.

[0079] Furthermore, similarly, the learning ability growth curves of trainees under different regions, fire types, and training frequencies can also be predicted.

[0080] S4. Initialize multiple candidate training scheme sequences and input them together with the historical training time series data and mine structure feature vector of the corresponding trainee into the first machine learning model to obtain the predicted score of the trainee under each candidate sequence; calculate the objective function value based on the predicted score, with the goal of maximizing the improvement of comprehensive ability, and determine the optimal candidate training scheme sequence through iterative search to obtain the trainee's simulated training plan sequence, including the following steps:

[0081] S4.1 Initialize multiple candidate training scheme sequences. Each candidate training scheme sequence consists of scheme items arranged in chronological order for a preset number of future training sessions. Each scheme item includes at least the mine area identifier, fire type, and training time interval.

[0082] S4.2 For each candidate training scheme sequence, the historical training time series data of the trainee is concatenated with the candidate training scheme sequence in the time dimension to form a joint input sequence. Among them, each time step of the historical training time series data includes the trainee's real-time score, and the position corresponding to the trainee's real-time score in each time step of the candidate training scheme sequence is a placeholder value. Each time step of the concatenated historical training time series data and candidate training scheme sequence also includes a type mask used to distinguish between the historical training time series data and the candidate training scheme sequence. For example, the type mask of the historical part can be set to 1, the type mask of the candidate part can be set to 0, and the placeholder value can be -1.

[0083] S4.3. Input the joint input sequence and the mine structure feature vector into the first machine learning model, and output the prediction score of the trainee for each fire type at each time step corresponding to the candidate training scheme sequence.

[0084] S4.4. Based on the predicted score, calculate the objective function value representing the improvement of the trainee's comprehensive ability under the candidate training scheme sequence; in one embodiment, the objective function is defined as:

[0085]

[0086] in, The improvement in the student's score in the k-th type of fire, as predicted by the first machine learning model after executing the candidate training sequence S (i.e., the difference between the value in the prediction head output matrix corresponding to the future R-th training position and the current score). Weights for fire type; The training cost and risk cost function; The balancing factor can be manually set by the instructor based on the availability of actual training resources. The system provides three default values: low, medium, and high.

[0087] S4.5. With the goal of maximizing the objective function value, the candidate training scheme sequence is updated through iterative search. The candidate training scheme sequence with the optimal objective function value under preset constraints is used as the customized training plan for the student simulation. The iterative search can employ genetic algorithms, Bayesian optimization algorithms, simulated annealing algorithms, etc.

[0088] S5. Based on the trainee simulation training plan sequence, generate control signals and send them to the corresponding disaster simulation module to control the disaster simulation module to automatically execute fire simulation.

[0089] Furthermore, the intelligent processing control module also includes the following steps:

[0090] During simulated fire training, the trainees' real-time scores are input into the first machine learning model for optimization calculation, and the subsequent parts of the trainees' simulated training plan sequence are adjusted based on the results of the optimization calculation.

[0091] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. An intelligent downhole simulated fire training system, characterized in that, It includes a data acquisition module, a human-computer interaction module, a disaster simulation module, and an intelligent processing and control module; The data acquisition module is used to synchronously collect trainee behavior data and environmental status data during the training process; The human-computer interaction module is used to visualize data and obtain user feedback information; The disaster simulation module has multiple simulations covering various fire types; The intelligent processing and control module is used to perform the following steps: S1. By using a pre-constructed knowledge graph that integrates expert rules and historical cases, the fire risk level of each structural feature of the mine is inferred based on the structural attribute information of the target mine, and an initial training scheme is generated based on the fire risk level. S2. During simulated fire training, the student behavior data and environmental status data collected in real time by the data acquisition module are input into the pre-trained scoring mapping model to generate real-time student scores. S3. Input the historical training time series data of each trainee into the pre-trained first machine learning model, and input the mine structure attribute information into the encoder obtained by the pre-trained first machine learning model as the conditional input of the first machine learning model, and predict the learning ability growth curve of the trainee under different fire types and training frequencies. S4. Initialize multiple candidate training scheme sequences and input them together with the historical training time series data and mine structure feature vector of the corresponding trainees into the first machine learning model to obtain the predicted score of the trainee under each candidate sequence; calculate the objective function value based on the predicted score, with the goal of maximizing the improvement of comprehensive ability, and determine the optimal candidate training scheme sequence through iterative search to obtain the trainee's simulated training plan sequence; S5. Based on the student simulation training plan sequence, generate control signals and send them to the corresponding disaster simulation module to control the disaster simulation module to automatically execute fire simulation.

2. The intelligent downhole simulated fire training system according to claim 1, characterized in that, The intelligent processing and control module also includes the following steps: Before each simulated fire training exercise, a global safety initialization is performed. After power-on, a global reset command with the highest priority is sent to all disaster simulation modules. After collecting reset completion confirmation signals returned by all disaster simulation modules, the system enters standby mode.

3. The intelligent downhole simulated fire training system according to claim 1, characterized in that, The intelligent processing and control module also includes the following steps: During the simulated fire training, the trainees' real-time scores are input into the first machine learning model for optimization calculation, and the subsequent parts of the trainees' simulated training plan sequence are adjusted based on the results of the optimization calculation.

4. The intelligent downhole simulated fire training system according to claim 1, characterized in that, S1 includes the following steps: The target mine's structural and static attribute information is obtained by receiving the structural and static attribute information of the target mine through the human-computer interaction module. Based on the structural attribute information of the target mine, in a knowledge graph consisting of the knowledge of the expert group and historical fire cases of mines similar to the target mine, with the probability and severity of the occurrence of corresponding fire types, the probability and severity of the occurrence of fire types in each region that match the structural attribute information of the target mine are retrieved by a rule engine or graph query language. The risk level of each fire type in each region of the target mine is calculated by weighting according to preset probability weights and severity weights, and the region-fire type-risk level sequence of the target mine is generated in descending order of risk level. The initial fire scenario training sequence is determined based on the order of the region-fire type-risk level sequence, and an initial training scheme is generated.

5. The intelligent downhole simulated fire training system according to claim 1, characterized in that, The pre-trained rating mapping model in S2 is pre-trained through the following steps: Using the student's historical behavior data sequence and corresponding environmental state data sequence as input, and the student's historical expert evaluation score as output, a deep neural network model is trained to obtain a score mapping model.

6. The intelligent downhole simulated fire training system according to claim 1, characterized in that, The first machine learning model and encoder in S3 are jointly trained through the following steps: For each trainee, the mine area, fire type, training time interval, and trainee's real-time score corresponding to each simulated fire training session are collected in chronological order to generate the trainee's historical training time sequence data; the training time interval is calculated based on the training time of two adjacent simulated fire training sessions; wherein, the training time interval of the first training session is set to a preset initial value; Historical training time series data is input into the sequence coding network of the first machine learning model to obtain a sequence coding vector; the sequence coding network is a neural network module used to extract temporal dependencies and contextual representations from a variable-length temporal input, and is used to compress the temporal patterns in the trainee's historical training time series data into a fixed-dimensional vector representation; The target mine structure attribute information is input into the encoder to obtain the target mine structure attribute feature vector; the encoder is jointly trained with the first machine learning model; The sequence encoding vector and the mine structural attribute feature vector are input into the gated fusion layer, concatenated, and then a gated vector is generated through a fully connected network. The gated vector is used to perform element-wise weighted combination of the sequence encoding vector and the mine structural attribute feature vector to obtain the fused feature vector. The fused feature vector is input into the prediction head network of the first machine learning model, and the prediction head network outputs the prediction score of each fire type in the trainee’s future fire simulation training for a preset number of times. The prediction head network is a fully connected network module that connects the fused features and the prediction target. The output dimension is determined by the number of fire types and the prediction step size. Using the real-time scores of the corresponding trainees in subsequent fire simulation training as the true labels, a weighted loss function between the predicted scores and the true labels is calculated, and the parameters of the sequence coding network, the gated fusion layer, the prediction head network, and the encoder of the first machine learning model are updated simultaneously through backpropagation.

7. The intelligent downhole simulated fire training system according to claim 6, characterized in that, The learning ability growth curve in S3 is composed of the predicted scores of each fire type output by the first machine learning model, arranged in the order of training for each student in the future preset number of training sessions; wherein, each predicted score represents the expected performance level of the student for the corresponding fire type under the corresponding training order.

8. The intelligent downhole simulated fire training system according to claim 1, characterized in that, S4 includes the following steps: Initialize multiple candidate training scheme sequences. Each candidate training scheme sequence consists of scheme items arranged in chronological order for a preset number of future training sessions. Each scheme item includes at least the mine area identifier, fire type, and training time interval. For each candidate training scheme sequence, the student's historical training time series data and the candidate training scheme sequence are concatenated in the time dimension to form a joint input sequence. Each time step of the historical training time series data includes the student's real-time score, and the position corresponding to the student's real-time score in each time step of the candidate training scheme sequence is a placeholder value. Each time step of the concatenated historical training time series data and the candidate training scheme sequence also includes a type mask used to distinguish between the historical training time series data and the candidate training scheme sequence. The joint input sequence and the mine structure feature vector are input into the first machine learning model, and the first machine learning model outputs the trainee's prediction score for each fire type at each time step corresponding to the candidate training scheme sequence. Based on the predicted score, calculate the objective function value that characterizes the improvement of the student's comprehensive ability under the candidate training scheme sequence; With the objective function value as the optimization objective, the candidate training scheme sequence is updated through iterative search. The candidate training scheme sequence with the optimal objective function value under preset constraints is used as the student simulation training customized plan.

Images