A method, device and equipment for fine tuning and distillation of a model in a coal mine underground

Abstract

This invention relates to the field of artificial intelligence technology in mining, and particularly to a method, apparatus, and equipment for fine-tuning and distillation of a coal mine underground model. The method includes: initial supervised fine-tuning of a teacher model and a target model using a precisely labeled dataset; performing multi-channel parallel inference on an unlabeled dataset using the initial teacher model and the target model, and constructing a high-quality distillation dataset after consistency verification and quality screening; fusing precisely labeled data, publicly available general data, and distillation data to form a comprehensive training set, and performing secondary supervised fine-tuning of the target model using multi-loss fusion; iteratively updating and quality-evaluating the distillation dataset based on the optimized target model, and performing enhanced fine-tuning using a thought chain inference template to obtain a final model adapted to edge devices. This invention effectively utilizes multi-source underground data and reduces labeling costs, significantly improving the model's adaptability, inference efficiency, and robustness in anomaly detection in complex underground environments, achieving a synergistic optimization of high accuracy and low resource consumption.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology in mining, and in particular to a method, apparatus and equipment for fine-tuning and distillation of underground coal mine models. Background Technology

[0002] Coal mine underground safety monitoring, as a core component of intelligent mine construction, is widely used in scenarios such as video analysis, target recognition, and anomaly early warning. With the development of deep learning technology, related technologies typically construct intelligent sensing systems adapted to edge devices through the collaborative operation of data acquisition, model fine-tuning, and knowledge distillation. Specifically, this system covers the entire process from multi-source data fusion processing to lightweight model deployment, including key steps such as precise sample training, unlabeled data mining, and general knowledge transfer, aiming to improve inference efficiency in complex environments.

[0003] However, existing model optimization methods directly adopt fine-tuning strategies for general scenarios without fully considering specific constraints such as insufficient underground lighting and dust interference. This results in low utilization of labeled data and insufficient value mining of unlabeled data. Current technologies lack mechanisms for efficiently utilizing unlabeled data, typically relying on a single model or inference path to generate pseudo-labels. This leads to low-quality pseudo-labels with high noise levels, negatively impacting subsequent model training. Furthermore, the lack of deep integration of reinforcement fine-tuning and logical guidance mechanisms results in poor inference robustness in complex tasks, making it difficult to effectively handle sudden anomalies. These technical limitations not only lead to poor model adaptability but also significantly increase data labeling costs and edge deployment difficulties, failing to meet the urgent need for real-time, high-precision monitoring in coal mines. Summary of the Invention

[0004] Therefore, the first objective of this invention is to propose a method for fine-tuning and distillation of coal mine underground models. This method, by constructing a multi-channel parallel inference and consistency verification mechanism, fully explores the potential value of unlabeled data, significantly improving the model's adaptability, inference efficiency, and robustness in anomaly detection in complex underground environments.

[0005] The second objective of this invention is to provide a fine-tuning and distillation device for underground coal mine models.

[0006] The third objective of this invention is to provide a coal mine underground model fine-tuning and distillation device.

[0007] To achieve the above objectives, a first aspect of the present invention proposes a method for fine-tuning and distillation of a coal mine underground model, comprising: acquiring a precisely labeled dataset, an unlabeled dataset, and a publicly available general dataset of a coal mine underground scene; using the precisely labeled dataset to perform initial supervised fine-tuning on a teacher model and a target model to obtain an initial teacher model and an initial target model; using the initial teacher model and the initial target model to perform multi-channel parallel inference on the unlabeled dataset to generate candidate labeling results, and performing consistency checks and quality screening on the candidate labeling results to construct a high-quality distillation dataset; wherein, the multi-channel parallel inference generates multiple candidate labeling results with diversity in content and inference paths by differentially configuring generation parameters, inference hints, inference structures, or contextual information, and performs ...; and performs quality screening on the candidate labeling results by differentially configuring generation parameters, inference hints, inference structures, or contextual information, and performs quality screening on the candidate labeling results. A dual-verification mechanism for scene logic matching compliance checks filters high-quality pseudo-labels; a comprehensive training dataset is formed by merging the precision-calibrated dataset, the publicly available general dataset, and the high-quality distillation dataset; the initial target model is then fine-tuned using a multi-loss fusion method with secondary supervision using this comprehensive training dataset to obtain an optimized target model; wherein, the multi-loss fusion comprehensively considers the supervision loss corresponding to the precision-calibrated data, the distillation loss corresponding to the distillation data, and the consistency loss of model prediction to collaboratively optimize model parameters; based on the optimized target model, the high-quality distillation dataset is iteratively updated and its quality is evaluated, and a reinforcement fine-tuning dataset containing logical guidance information is constructed using the updated data to reinforce and fine-tune the optimized target model, resulting in a final model adapted to underground coal mine edge equipment.

[0008] To achieve the above objectives, a second aspect of the present invention proposes a coal mine underground model fine-tuning and distillation device, comprising: an initial fine-tuning module, used to acquire a finely labeled dataset, an unlabeled dataset, and a publicly available general dataset of a coal mine underground scene, and to perform initial supervised fine-tuning of a teacher model and a target model using the finely labeled dataset to obtain an initial teacher model and an initial target model; and a distillation data construction module, used to perform multi-channel parallel inference on the unlabeled dataset using the initial teacher model and the initial target model to generate candidate labeling results, and to perform consistency verification and quality screening on the candidate labeling results to construct a high-quality distillation dataset; wherein, the multi-channel parallel inference generates multiple candidate labeling results with diversity in content and inference path by differentially configuring generation parameters, inference hints, inference structures, or contextual information, and performs consistency checks and quality screening on the candidate labeling results to construct a high-quality distillation dataset; wherein, the multi-channel parallel inference generates multiple candidate labeling results with diversity in content and inference path by differentially configuring generation parameters, inference hints, inference structures, or contextual information, and performs consistency checks and quality screening on the candidate labeling results. A dual verification mechanism for compliance checks on the logic matching degree of the next scene filters out high-quality pseudo-labels; a secondary fine-tuning module is used to merge the precision-calibrated dataset, the publicly available general dataset, and the high-quality distillation dataset to form a comprehensive training dataset, and to perform secondary supervised fine-tuning of the initial target model using the comprehensive training dataset with multi-loss fusion to obtain an optimized target model; wherein, the multi-loss fusion comprehensively considers the supervision loss corresponding to the precision-calibrated data, the distillation loss corresponding to the distillation data, and the consistency loss of model prediction to collaboratively optimize model parameters; a reinforcement fine-tuning module is used to iteratively update and evaluate the quality of the high-quality distillation dataset based on the optimized target model, and to construct a reinforcement fine-tuning dataset containing logical guidance information using the updated data, and to perform reinforcement fine-tuning of the optimized target model to obtain the final model adapted to the edge equipment in underground coal mines.

[0009] To achieve the above objectives, a third aspect of the present invention provides a coal mine underground model fine-tuning and distillation device, comprising: a processor; a memory storing executable instructions; wherein when the processor executes the executable instructions, it implements the method described in the first aspect of the present invention.

[0010] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0011] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0012] Figure 1 This is a flowchart illustrating the implementation of a coal mine underground model fine-tuning and distillation method provided by the present invention. Figure 2 This is a structural block diagram of a coal mine underground model fine-tuning and distillation device provided in an embodiment of the present invention. Detailed Implementation

[0013] The core of this invention is to provide a method, apparatus, and equipment for fine-tuning and distillation of underground coal mine models. This solves the problems of poor model adaptability and low data utilization in underground coal mine scenarios, significantly improving the detection accuracy and inference efficiency of the model in complex environments. Simultaneously, through lightweight distillation and iterative optimization mechanisms, it reduces data annotation costs and the deployment threshold for edge devices, achieving synergistic optimization of high accuracy and low resource consumption.

[0014] To enable those skilled in the art to better understand the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] Please refer to Figure 1. Figure 1 The flowchart illustrates the implementation of a coal mine underground model fine-tuning and distillation method provided by this invention; the specific operation steps are as follows: S101: Obtain the precise labeled dataset, unlabeled dataset, and publicly available general dataset for the underground coal mine scenario. Use the precise labeled dataset to perform initial supervised fine-tuning on the teacher model and the target model to obtain the initial teacher model and the initial target model. Specifically, multi-source heterogeneous data from underground coal mine scenarios are acquired and a hierarchical data system is constructed. Then, the finely labeled data within this system is used to perform initial supervised fine-tuning on the pre-set teacher and target models, endowing the models with basic underground task processing capabilities. The core of this step lies in injecting prior knowledge of specific underground coal mine scenarios into the model parameters through a supervised learning mechanism, establishing the initial feature extraction and recognition benchmark for key underground task objectives. During this process, the finely labeled data serves as a supervisory signal source, used to calculate the deviation between the model's predicted distribution and the true labels, thereby driving iterative updates to the model parameters and rapidly converging to a solution space suitable for the underground environment in the early stages of training. The teacher and target models are trained simultaneously at this stage, aiming to lay a unified feature representation foundation for subsequent distillation and optimization processes.

[0016] This step, by constructing a hierarchical data system and using precisely labeled data for initial supervised fine-tuning, effectively addresses the problem of poor adaptability of general-purpose models in the complex environment of underground coal mines, significantly reducing the difficulty of cold-starting model training. Simultaneously, by synchronously initializing the teacher and target models, not only is the model's underlying general feature extraction capability preserved, but a reliable foundation for generating pseudo-labels is also provided for subsequent high-quality distillation using unlabeled data, thereby improving the efficiency and stability of the overall training process.

[0017] S102: Using the initial teacher model and the initial target model, perform multi-channel parallel inference on the unlabeled dataset to generate candidate labeling results, and perform consistency verification and quality screening on the candidate labeling results to construct a high-quality distillation dataset; wherein, the multi-channel parallel inference generates multiple candidate labeling results with diversity in content and inference path by configuring generation parameters, inference hints, inference structure or context information in a differentiated manner, and filters high-quality pseudo-labels through a dual verification mechanism including consistency check and underground scene logic matching degree compliance check; This step aims to address the challenge of directly utilizing unlabeled data from underground coal mines. By constructing a pseudo-label generation and filtering mechanism based on multi-model collaboration, the original unlabeled data is transformed into highly reliable training resources. The core of this approach lies in leveraging an initial teacher model and an initial target model, both already capable of basic task processing, to perform multi-channel parallel inference on the unlabeled dataset. By differentiating inference parameters, hint strategies, or structural paths, the model's inference diversity across the multi-dimensional feature space is stimulated, generating multiple candidate labeling results covering varying confidence levels and semantic understandings. Subsequently, a consistency verification and quality screening mechanism is introduced to logically compare and analyze the candidate results generated from multiple channels. Abnormal labels that are contradictory, have low confidence, or do not conform to prior knowledge of the underground scenario are eliminated, retaining only highly consistent and high-quality labeled data, thereby constructing a high-quality distillation dataset.

[0018] This step fully explores the potential value of unlabeled data through multi-channel parallel inference, and effectively suppresses pseudo-label noise by combining strict consistency verification, significantly improving the accuracy and reliability of distillation data, providing stable and rich data support for subsequent model training, while reducing the dependence on manually labeled data and improving data utilization efficiency.

[0019] S103: The precise calibration dataset, the publicly available general dataset, and the high-quality distillation dataset are merged to form a comprehensive training dataset. The initial target model is then fine-tuned using the comprehensive training dataset with secondary supervision through multi-loss fusion to obtain the optimized target model. The multi-loss fusion comprehensively considers the supervision loss corresponding to the precise calibration data, the distillation loss corresponding to the distillation data, and the consistency loss of the model prediction to collaboratively optimize the model parameters. This step aims to address the technical problems of insufficient model generalization ability and poor scenario adaptability caused by a single data source by constructing a multi-dimensional data fusion mechanism and a multi-objective optimization strategy. Its core lies in organically integrating multiple datasets with different feature distributions and annotation attributes to form a comprehensive training dataset that can fully cover the complex working conditions of underground coal mines. This dataset then drives the target model to perform deep parameter iteration and feature reconstruction. In this process, a multi-loss fusion optimization mechanism is adopted. By constructing a composite loss function that includes supervision signals, distillation signals, and consistency constraints, the update direction of model parameters is guided collaboratively. This allows the model to retain its general feature extraction capabilities while significantly enhancing its perception accuracy and inference robustness for specific underground scenarios. This multi-loss fusion mechanism effectively avoids the risk of overfitting by dynamically balancing the contribution weights of various data types to model training, achieving a synergistic improvement in model accuracy and generalization performance.

[0020] Through the aforementioned multi-source data fusion and multi-loss collaborative optimization methods, this step can fully explore the potential value of unlabeled and publicly available data, significantly improve the model's adaptability and detection accuracy in complex environments such as low illumination and high dust in coal mines, and at the same time ensure the efficient operation of the model on edge devices with limited computing power, achieving the best balance between training effect and deployment cost.

[0021] S104: Based on the optimized target model, iteratively update and evaluate the quality of the high-quality distillation dataset, and use the updated data to construct an enhanced fine-tuning dataset containing logical guidance information, and enhance and fine-tune the optimized target model to obtain the final model adapted to the underground edge equipment of the coal mine.

[0022] This step aims to further enhance the model's inference robustness in complex scenarios and generate a final model adapted to edge devices through iterative optimization mechanisms and logical guidance strategies. Its core lies in constructing a closed loop of data and model co-evolution. The optimized target model is used to iteratively update the existing distillation dataset with pseudo-labels, and a quality evaluation mechanism is introduced to filter high-value samples, thereby dynamically improving the accuracy and scenario adaptability of the training data. Building upon this, the method constructs a reinforced fine-tuning dataset by injecting logical guidance information, integrating domain knowledge or inference rules into the training process to perform targeted reinforcement and fine-tuning of the model, thereby enhancing the model's logical adherence to key task nodes and the stability of anomaly detection.

[0023] Through the aforementioned iterative update and quality evaluation mechanism, the accumulated errors in distillation data can be effectively corrected, significantly improving data quality and model generalization ability. At the same time, the introduction of enhanced fine-tuning with logical guidance information not only enhances the model's logical reasoning depth and anomaly warning robustness in complex downhole environments, but also ensures that the final model meets the deployment requirements of edge computing devices while maintaining high accuracy, achieving a synergistic improvement in accuracy, efficiency, and scenario adaptability.

[0024] Based on the above embodiments, in some embodiments, the step of acquiring a precisely labeled dataset, an unlabeled dataset, and a publicly available dataset of a coal mine underground scenario, and using the precisely labeled dataset to perform initial supervised fine-tuning of the teacher model and the target model to obtain an initial teacher model and an initial target model includes: Video data is collected and preprocessed by downhole monitoring equipment. Candidate data containing key task objectives are extracted and labeled to form a finely labeled dataset. The remaining unlabeled data is used to form an unlabeled dataset. Publicly available general data that is adapted to the downhole scenario is selected to form a publicly available general dataset. A model whose feature extraction capability meets the accuracy requirements is selected as the teacher model, and a lightweight model is selected as the target model. The teacher model and the target model are simultaneously subjected to initial supervised fine-tuning based on the precision dataset, and some low-level parameters are frozen during the training process. Set training hyperparameters, use a loss function to calculate the deviation between the model prediction results and the precise label to guide the model parameter update, and save the model parameters with basic downhole task processing capabilities through iterative evaluation to obtain the initial teacher model and the initial target model.

[0025] In other embodiments, the input source is the raw video stream collected by monitoring equipment deployed in key underground coal mine scenarios. Processing actions include image enhancement and denoising preprocessing tailored to the low-light, high-dust environment of underground mines. Subsequently, candidate data segments containing key task targets such as personnel and equipment are extracted and manually precisely labeled. The output is a precisely labeled dataset. Simultaneously, the remaining unlabeled data is organized into an unlabeled dataset, and publicly available general data adapted to the mine's geology and operational scenarios is selected to form a publicly available general dataset. Alternatively, the input source consists of a selected deep learning model with strong feature extraction capabilities and high inference accuracy as the teacher model, a lightweight deep learning model as the target model, and the aforementioned precisely labeled dataset. Processing actions include freezing some low-level parameters of both models during training to protect general feature extraction capabilities and reduce computational costs, setting hyperparameters for learning rate, batch size, and iteration count, and using the cross-entropy loss function as the core loss function.

[0026] In one specific embodiment: a deep learning model with strong feature extraction capabilities and high inference accuracy (such as Qwen3-vl-32B) is selected as the teacher model, mainly used to provide high-quality pseudo-labels for unlabeled data; a lightweight deep learning model (such as Qwen3-vl-2B) is selected as the target model. This model has fewer parameters and lower computational overhead, and is designed specifically for subsequent distillation training and deployment on edge computing devices (such as embedded industrial control computers and smart cameras) in coal mines. Based on the constructed precise labeled dataset, both the teacher model and the target model undergo initial supervised fine-tuning simultaneously. During training, some low-level parameters of the model (such as the first few layers of a convolutional neural network or the first few encoding layers of a Transformer) are frozen. This reduces training computation costs and effectively protects the general feature extraction capabilities (such as edge and texture detection) already learned by the model's underlying layers. Specifically, the learning rate is set to 0.001, the batch size to 32, and the number of iterations to 50 rounds. The cross-entropy loss function is used as the core loss function for initial supervised fine-tuning.

[0027] Specifically: the formula for calculating the cross-entropy loss function is as follows: ,in This refers to the number of training samples in a single session. For the precise labeling of the samples, The model predicts the probability distribution. In each training iteration, the precise labels of the samples in the current batch and the predicted probability distribution output by the model are obtained. These are then substituted into the above formula to perform logarithmic operations and weighted summation to calculate the supervised loss value for the representation bias. Based on this loss value, the backpropagation algorithm is executed to calculate the gradient information of the parameters of each layer of the model. The gradient information is then used to update the network weight parameters of the unfrozen parts of the teacher model and the target model. Through iterative evaluation, when the model's performance on the validation set reaches the preset standard and tends to stabilize over multiple rounds, the parameters of the initial teacher model and the initial target model, which have basic downhole task processing capabilities, are saved.

[0028] It should be noted that 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 technical features indicated. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly defined. This embodiment effectively reduces the computational resource consumption in the initial stage by freezing the underlying parameters and employing a refined training strategy using the cross-entropy loss function. Simultaneously, it rapidly endows the model with basic downhole scene feature extraction and target recognition capabilities using precisely labeled data, laying a solid model foundation for the high-quality generation of subsequent distillation data.

[0029] Based on the above embodiments, in some embodiments, using a loss function to calculate the deviation between the model prediction result and the precise label to guide model parameter updates includes: In each training iteration, the sample labels and the predicted probability distribution of the model output are obtained for the current batch of samples. The sample labels and the predicted probability distribution are substituted into the loss function to perform the calculation, and the supervision loss value that represents the deviation between the model prediction result and the label is calculated. Based on the supervised loss value, the backpropagation algorithm is executed to calculate the gradient information of the parameters of each layer of the model, and the gradient information is used to update the network weight parameters of the unfrozen parts of the teacher model and the target model; The updated network weight parameters are used in the next training iteration until the preset number of iterations or model convergence conditions are reached.

[0030] In other embodiments, the step of using the cross-entropy loss function as the core loss function for initial supervised fine-tuning to calculate the deviation between the model prediction results and the precise labels to guide model parameter updates includes: obtaining the precise labels of the samples in the current batch during each training iteration. and the predicted probability distribution of the model output The sample is labeled with precision. and the predicted probability distribution Substitute into the formula The algorithm performs logarithmic and weighted summation operations to calculate the supervised loss value of the bias between the prediction results of the representation model and the precise label. Based on the calculated supervision loss value The backpropagation algorithm is executed to calculate the gradient information of the parameters of each layer of the model. The gradient information is used to update the network weight parameters of the unfrozen parts of the teacher model and the target model. The updated network weight parameters are used for the next round of training iterations until the preset number of iterations or the model convergence condition is reached, thus completing the parameter optimization process based on the cross-entropy loss function.

[0031] In one specific embodiment: the learning rate is set to 0.001, the batch size to 32, and the number of iterations to 50 rounds. The cross-entropy loss function is used as the core loss function for initial supervised fine-tuning. In each training iteration, the precise labels of the samples in the current batch and the predicted probability distribution of the model output are obtained, substituted into the cross-entropy loss function formula, and the supervised loss value is calculated. Based on this loss value, the backpropagation algorithm is executed to calculate the gradient information of the parameters of each layer of the model, and the gradient information is used to update the network weight parameters of the unfrozen parts in the teacher model and the target model.

[0032] Specifically: The updated network weight parameters are used in the next round of training iterations until the preset number of iterations or model convergence conditions are reached, and the initial teacher model and initial target model parameters with basic downhole task processing capabilities are saved.

[0033] It should be noted that this implementation method effectively reduces the computational resource consumption in the initial stage by freezing the underlying parameters and adopting a refined training strategy using the cross-entropy loss function. At the same time, it utilizes the finely labeled data to quickly endow the model with basic downhole scene feature extraction and target recognition capabilities, laying a solid model foundation for the high-quality generation of subsequent distillation data.

[0034] Based on the above embodiments, in some embodiments, the process of using the initial teacher model and the initial target model to perform multi-channel parallel inference on the unlabeled dataset to generate candidate labeling results, and performing consistency verification and quality screening on the candidate labeling results to construct a high-quality distillation dataset; wherein, the multi-channel parallel inference generates multiple candidate labeling results with diversity in content and inference path by configuring generation parameters, inference hints, inference structure or context information differently, and filters high-quality pseudo-labels through a dual verification mechanism including consistency checks and compliance checks of the logic matching degree of the well scene, including: The same unlabeled input data from the unlabeled dataset is simultaneously input into an inference system configured with multiple parallel generation channels. Each parallel generation channel performs independent inference on the unlabeled input data through differentiated configurations of generation parameters, inference hints, inference structure, or context information, generating multiple candidate labeling results with diversity in content and inference path. The multiple candidate annotation results are input into the quality assurance module, and a dual verification mechanism based on consistency constraints is executed to eliminate low-quality annotations that do not meet the quality standards. The dual verification mechanism includes a consistency check by comparing the multiple candidate annotation results with each other and a compliance check on the matching degree between the annotation results and the downhole scene logic. Candidate annotations that pass the dual verification mechanism are selected as high-quality pseudo-labels. These high-quality pseudo-labels are then associated with and stored with the unlabeled input data to construct a high-quality distillation dataset.

[0035] In other embodiments, the same unlabeled input data from the unlabeled dataset is used as a unified input source and simultaneously distributed to an inference system configured with multiple parallel generation channels. Each parallel generation channel operates independently based on a differentiated configuration strategy. Specifically, by adjusting generation parameters such as the temperature coefficient and the number of top kernel samples for each channel, or by loading differentiated inference prompt templates designed for different underground task scenarios, or by using different attention mask structures and context window lengths, independent inference is performed on the same unlabeled input data, thereby generating multiple candidate annotation results with significant diversity in content description and logical inference paths. Subsequently, the multiple candidate annotation results generated above are sent as input sources to the quality assurance module to perform a dual verification mechanism based on consistency constraints. The first verification of this mechanism is a consistency check, which calculates the semantic similarity or edit distance between multiple candidate annotation results and removes outlier annotations that deviate from other results by more than a preset threshold. The second verification is a compliance check, which matches the remaining candidate annotation results with a preset logical rule base for underground coal mine scenarios to verify whether they conform to the operating rules of underground equipment, personnel behavior norms, and environmental physical characteristics, and removes low-quality annotations that have logical conflicts or violate common sense in the scenario.

[0036] In one specific embodiment: multiple parallel generation channels are set up, each channel is configured differently in terms of generation parameters or context information. After generating candidate labels, they are filtered through a dual verification mechanism to finally obtain a high-quality distillation dataset S_dist. The same unlabeled input data from the unlabeled dataset is simultaneously input into an inference system configured with multiple parallel generation channels. Each parallel generation channel performs independent inference on the same unlabeled input data by different configurations of generation parameters (such as different settings of the temperature coefficient), inference hints (different expressions of Prompt Engineering), inference structures (such as different decoding strategies Beam Search and Sampling), or context information, thereby generating multiple candidate label results with diversity in content and inference paths, fully expanding the model's inference space.

[0037] Specifically, after a dual verification mechanism, the candidate annotations that pass the verification are determined as high-quality pseudo-labels and used as the final output. Finally, these high-quality pseudo-labels are stored with key-value pairs with the corresponding unlabeled input data to form a high-quality distillation dataset containing stable and reliable annotation information. This dataset directly serves as a core component of the comprehensive training dataset in the subsequent step S103, providing data support for secondary supervised fine-tuning of multi-loss fusion.

[0038] It should be noted that this implementation method effectively expands the inference space of unlabeled data through multi-channel parallel inference and differentiated configuration strategies, avoiding deviations caused by a single inference path. Combined with a dual verification mechanism of consistency check and scenario logic compliance check, it significantly improves the accuracy and reliability of pseudo-labels, ensures the quality of the distillation dataset, and thus enhances the generalization ability and robustness of the target model in the complex environment of underground coal mines.

[0039] Based on the above embodiments, in some embodiments, the fusion of the precision-calibrated dataset, the publicly available general dataset, and the high-quality distillation dataset to form a comprehensive training dataset, and the use of the comprehensive training dataset to perform secondary supervised fine-tuning of the initial target model using multi-loss fusion, to obtain an optimized target model; wherein, the multi-loss fusion comprehensively considers the supervision loss corresponding to the precision-calibrated data, the distillation loss corresponding to the distillation data, and the consistency loss of model prediction, in order to collaboratively optimize the model parameters, including: The precision dataset, the publicly available dataset, and the high-quality distillation dataset are mixed and spliced ​​together according to a preset ratio to form a comprehensive training dataset containing multi-source heterogeneous data. Remove some of the frozen parameters from the initial target model, and perform secondary supervised fine-tuning training on the unfrozen initial target model using the comprehensive training dataset; During training, a fusion loss function is applied to comprehensively consider the supervision loss corresponding to the labeled data, the distillation loss corresponding to the distillation data, and the consistency loss of the model prediction, so as to optimize the model training effect. The model performance is evaluated through periodic iterations. When the model performance reaches the preset standard and tends to stabilize after several consecutive rounds, the secondary supervised fine-tuning is stopped, and the optimized target model parameters are saved as the optimized target model.

[0040] In other embodiments, a standardized dataset, a publicly available general dataset, and a high-quality distillation dataset are used as input sources. Data is read and randomly shuffled according to a preset mixing ratio, and a comprehensive training dataset containing multi-source heterogeneous data is constructed through a concatenation operation. This dataset serves as the input basis for subsequent model training. Next, the initial target model, after initial supervised fine-tuning, is loaded, and some low-level network parameters are unfrozen, making them trainable parameters to optimize the model's feature extraction capability and adaptability to complex underground coal mine scenarios. Secondary supervised fine-tuning training is then performed on the unfrozen initial target model using the comprehensive training dataset.

[0041] In one specific embodiment: the precision-calibrated dataset, the publicly available general dataset, and the high-quality distillation dataset can be directly fused to form a comprehensive training dataset. During the secondary supervised fine-tuning stage, some frozen parameters are unfrozen, and a fusion loss function is employed. As the optimization objective, each item corresponds to the supervision loss of the precise data, the distillation loss of the distillation data, and the prediction consistency loss, respectively. The optimized target model parameters are saved through iterative evaluation.

[0042] Specifically: In each training iteration, the supervised loss generated by the precisely labeled data is calculated. Distillation losses generated from distillation data and consistency loss between model predictions Obtain the preset loss weight coefficients. , , Multiply the three losses mentioned above by their corresponding weighting coefficients, perform addition, and then substitute them into the formula. Calculate the total loss value Based on this total loss value The gradient backpropagation operation is performed to automatically calculate the gradient update of each trainable parameter in the initial target model. The optimizer then adjusts the model network weights based on the gradient updates to minimize the total loss. The adjusted model network weights are used as input for the next iteration of training, forming a closed-loop optimization process of "loss calculation - gradient backpropagation - weight update". During this process, the model performance is periodically evaluated using a validation set. When the model performance metrics reach a preset standard and tend to stabilize over multiple iterations, the stopping condition is met, the secondary supervised fine-tuning training is terminated, and the current model network weights are saved as the optimized target model parameters.

[0043] It should be noted that this implementation method effectively improves the feature adaptation capability and inference accuracy of the target model in the complex environment of underground coal mines through multi-source data fusion and multi-loss function collaborative optimization. At the same time, it avoids overfitting by using a dynamic stopping mechanism, ensuring the efficiency and stability of model deployment.

[0044] Based on the above embodiments, in some embodiments, the application of a fusion loss function during training, which comprehensively considers the supervision loss corresponding to the labeled data, the distillation loss corresponding to the distillation data, and the consistency loss of the model prediction, to optimize the model training effect includes: In each training iteration, the supervision loss generated by the finely labeled data, the distillation loss generated by the distillation data, and the consistency loss between the model prediction results are calculated separately. Obtain the preset loss weight coefficients, multiply the supervision loss, the distillation loss and the consistency loss by their respective loss weight coefficients, perform an addition operation, and substitute them into the fusion loss function to calculate the total loss value; Based on the total loss value, perform gradient backpropagation to calculate the gradient update amount of each trainable parameter in the initial target model, and use the gradient update amount to adjust the model network weights to minimize the total loss value. The adjusted model network weights are used in the next iteration of training until the preset stopping condition is met.

[0045] In other embodiments, the formula for calculating the fusion loss function is as follows: in For loss weighting coefficients, To monitor losses, For distillation losses, This is the consistency loss. The supervised loss generated by the precisely labeled data is calculated in each training iteration. Distillation losses generated from distillation data and consistency loss between model predictions ; Obtain the preset loss weight coefficients The monitoring loss The distillation loss and the aforementioned consistency loss After multiplying by the corresponding loss weight coefficients, perform the addition operation and substitute into the formula. Calculate the total loss value Based on the total loss value Perform gradient backpropagation to calculate the gradient update of each trainable parameter in the initial target model, and adjust the model network weights using the gradient update to minimize the total loss. The adjusted model network weights are used in the next iteration of training until the preset stopping condition is met, thereby achieving collaborative optimization of model parameters under multi-loss fusion.

[0046] In one specific embodiment: α=0.4, β=0.4, γ=0.2; To monitor losses, This is the distillation loss (usually calculated as the KL divergence between the output distribution of the target model and the output distribution of the teacher model). The consistency loss is used to constrain the model's prediction consistency for the same input under different augmented views. In each training iteration, the above three loss values ​​are calculated separately, multiplied by their corresponding weight coefficients, and then summed to obtain the total loss value. Based on the total loss value, perform gradient backpropagation to calculate the gradient update amount of each trainable parameter in the initial target model, and use the update amount to adjust the model network weights to minimize the total loss value.

[0047] Specifically: The adjusted model network weights are used in the next iteration of training until a preset stopping condition is met. The model performance is evaluated through periodic iterations. When the model performance reaches a preset standard and tends to stabilize over multiple rounds, the secondary supervised fine-tuning is stopped, and the optimized target model parameters are saved.

[0048] It should be noted that this implementation method effectively improves the feature adaptation capability and inference accuracy of the target model in the complex environment of underground coal mines through multi-source data fusion and multi-loss function collaborative optimization. At the same time, it avoids overfitting by using a dynamic stopping mechanism, ensuring the efficiency and stability of model deployment.

[0049] Based on the above embodiments, in some embodiments, the iterative updating and quality evaluation of the high-quality distillation dataset based on the optimized target model, and the construction of a reinforced fine-tuning dataset containing logical guidance information using the updated data, to reinforce and fine-tune the optimized target model to obtain the final model adapted to underground coal mine edge equipment includes: The optimized target model is used to re-perform inference labeling on the data in the high-quality distillation dataset to update the pseudo-labels; An evaluation model is introduced to score and verify the updated distillation dataset. The scoring and verification covers the dimensions of labeling accuracy, sample fit and content completeness, and filters out problematic data with labeling errors or substandard quality. The problematic data is re-input into the multi-channel parallel inference process for parallel inference and quality verification again, forming an iterative closed loop until the proportion of problematic data drops below a preset threshold or reaches a preset number of iterations. Based on the needs of underground coal mine monitoring tasks, a thought chain reasoning template adapted to the underground scenario is constructed. Samples from the precision-standard dataset and the iteratively updated distillation dataset are selected and added to the thought chain reasoning template to construct an enhanced fine-tuning dataset containing logical guidance information. The target model after iterative optimization is subjected to reinforcement fine-tuning using the aforementioned reinforcement fine-tuning dataset to obtain the final model adapted to underground edge equipment in coal mines.

[0050] In other embodiments, the iterative updating and quality evaluation of the high-quality distillation dataset based on the optimized target model, and the construction of a reinforced fine-tuning dataset containing logical guidance information using the updated data, followed by further refinement of the optimized target model to obtain a final model adapted to underground coal mine edge equipment, includes: re-executing inference labeling on the data in the high-quality distillation dataset using the optimized target model to update pseudo-labels and improve labeling accuracy; introducing a specialized evaluation model to comprehensively score and verify the updated distillation dataset, the scoring and verification covering dimensions of labeling accuracy, sample fit, and content completeness, filtering out problematic data with labeling errors or substandard quality; and re-inputting the problematic data into a multi-channel parallel push... The process involves parallel reasoning and quality verification, forming an iterative closed loop of model optimization, data updating, and retraining, until the proportion of problematic data drops below a preset threshold or the preset number of iterations is reached. A thought chain reasoning template adapted to underground scenarios is constructed based on the actual needs of various monitoring tasks in coal mines. The core process and key nodes of model reasoning are clarified, and typical samples from the precision dataset and the iteratively updated distillation dataset are added to the thought chain reasoning template to construct a reinforced fine-tuning dataset containing logical guidance information. The reinforced fine-tuning dataset is used to perform reinforcement fine-tuning on the iteratively optimized target model, focusing on optimizing the model's logical reasoning ability and robustness in recognizing complex underground scenarios and sudden anomalies, resulting in a final model adapted to underground coal mine edge equipment.

[0051] In one specific embodiment: the optimized target model obtained in step S103 is used as the inference engine, and the input source is the high-quality distillation dataset constructed in step S102. For unlabeled samples in the dataset, the processing involves re-inferring and labeling the input samples using the optimized target model, generating updated pseudo-labels to correct initial labeling biases. The output is an updated distillation dataset with improved labeling accuracy. Next, a pre-trained evaluation model is introduced. Each sample in the updated distillation dataset is sequentially input into this model. Processing includes evaluating the matching degree between pseudo-labels and the sample's true features based on labeling accuracy; evaluating whether the sample conforms to the lighting, dust, and composition characteristics of a specific underground coal mine scenario based on sample suitability; and evaluating the lack of information on key task objectives in the sample based on content completeness. The evaluation results from these three dimensions are then combined to generate a comprehensive quality score for each sample. This score is compared with a preset quality pass threshold, and the output is a set of problematic data judged as mislabeled or substandard. These data are then labeled and separated from the current dataset. Subsequently, the separated problematic data is transmitted as re-input data to the input end of the multi-channel parallel inference process, triggering a new round of parallel inference and quality verification, forming an iterative closed loop of "model optimization - data update - retraining". This process is repeated until the proportion of problematic data drops below a preset threshold or reaches a preset number of iterations, and the output is the iteratively updated distillation dataset that meets the quality standards. Based on this, a thought chain inference template adapted to the underground scenario is constructed in combination with the actual needs of various monitoring tasks in coal mines, clarifying the core process and key nodes of model inference, with the input source being the precision dataset from step S101. The processing action involves adding the aforementioned thought chain reasoning template to the selected typical samples in the updated distillation dataset. The output is a reinforced fine-tuning dataset containing logical guidance information. Finally, this reinforced fine-tuning dataset is used to perform reinforcement fine-tuning on the iteratively optimized target model, focusing on optimizing the model's logical reasoning ability and robustness in recognizing complex underground scenarios and sudden anomalies. The final output is a model adapted to underground edge equipment in coal mines.

[0052] Specifically: The introduction of a specialized evaluation model to comprehensively score and verify the updated distillation dataset. This scoring and verification covers dimensions of annotation accuracy, sample fit, and content completeness. Screening out problematic data with annotation errors or substandard quality involves: inputting each sample in the updated distillation dataset into a pre-trained evaluation model. This model evaluates the degree of matching between pseudo-labels and the true features of the samples based on annotation accuracy; assesses whether the samples conform to the lighting, dust, and composition characteristics of a specific underground coal mine scenario based on sample fit; and evaluates the lack of information regarding key task objectives in the samples based on content completeness. The evaluation results of the annotation accuracy index, the sample fit index, and the content completeness index are combined to generate a comprehensive quality score for each sample. The comprehensive quality score is compared with a preset quality pass threshold. Samples below the quality pass threshold are identified as problematic data with annotation errors or substandard quality, and these problematic data are marked and separated from the current distillation dataset. The marked and separated problematic data is used as re-input data and sent to the input end of the multi-channel parallel inference process to trigger a new round of parallel inference and quality verification process, so as to achieve dynamic cleaning and quality improvement of distillation data.

[0053] It should be noted that this implementation method constructs an evaluation model to perform multi-dimensional scoring and verification and dynamic cleaning of distillation data, effectively eliminating low-quality pseudo-label data and significantly improving the reliability of distillation data. At the same time, it introduces a thought chain reasoning template to construct an enhanced fine-tuning dataset, which enhances the model's logical reasoning ability and anomaly identification robustness in the complex environment of underground coal mines, ensuring the high-precision operation of the final model on edge devices.

[0054] Please refer to Figure 2 , Figure 2 A structural block diagram of a coal mine underground model fine-tuning and distillation device provided in this embodiment of the invention; the specific device may include: The initial fine-tuning module 100 is used to acquire a finely labeled dataset, an unlabeled dataset, and a publicly available general dataset for a coal mine underground scenario. The finely labeled dataset is used to perform initial supervised fine-tuning on the teacher model and the target model to obtain the initial teacher model and the initial target model. The distillation data construction module 200 is used to perform multi-channel parallel inference on the unlabeled dataset using the initial teacher model and the initial target model to generate candidate labeling results, and to perform consistency verification and quality screening on the candidate labeling results to construct a high-quality distillation dataset. The multi-channel parallel inference generates multiple candidate labeling results with diversity in content and inference path by configuring generation parameters, inference hints, inference structure or context information in a differentiated manner, and filters high-quality pseudo-labels through a dual verification mechanism including consistency check and underground scene logic matching degree compliance check. The secondary fine-tuning module 300 is used to fuse the precision-calibrated dataset, the publicly available general dataset, and the high-quality distillation dataset to form a comprehensive training dataset. The comprehensive training dataset is used to perform secondary supervised fine-tuning of the initial target model using multi-loss fusion to obtain an optimized target model. The multi-loss fusion comprehensively considers the supervision loss corresponding to the precision-calibrated data, the distillation loss corresponding to the distillation data, and the consistency loss of the model prediction to collaboratively optimize the model parameters. The enhancement and fine-tuning module 400 is used to iteratively update and evaluate the quality of the high-quality distillation dataset based on the optimized target model, and to construct an enhancement and fine-tuning dataset containing logical guidance information using the updated data, thereby enhancing and fine-tuning the optimized target model to obtain the final model adapted to the underground edge equipment of the coal mine.

[0055] In some embodiments, the distillation data construction module includes: The annotation candidate unit is used to simultaneously input the same unlabeled input data from the unlabeled dataset into an inference system configured with multiple parallel generation channels. Each of the parallel generation channels performs independent inference on the unlabeled input data through differentiated configurations of generation parameters, inference hints, inference structures, or contextual information, generating multiple candidate annotation results with diversity in content and inference paths. The annotation filtering unit is used to input the multiple candidate annotation results into the quality assurance module and execute a dual verification mechanism based on consistency constraints to eliminate low-quality annotations that do not meet the quality standards. The dual verification mechanism includes a consistency check by comparing the multiple candidate annotation results with each other and a compliance check by checking the matching degree between the annotation results and the downhole scene logic. A high-quality distillation dataset construction unit is used to filter candidate annotation results that pass the dual verification mechanism as high-quality pseudo-labels, associate and store the high-quality pseudo-labels with the unlabeled input data, and construct a high-quality distillation dataset.

[0056] The coal mine underground model fine-tuning and distillation device of this embodiment is used to implement the aforementioned coal mine underground model fine-tuning and distillation method. Therefore, the specific implementation of the coal mine underground model fine-tuning and distillation device can be found in the embodiment section of the coal mine underground model fine-tuning and distillation method above. For example, the initial fine-tuning module 100, the distillation data construction module 200, the secondary fine-tuning module 300, and the enhanced fine-tuning module 400 are respectively used to implement steps S101, S102, S103, and S104 in the above-mentioned coal mine underground model fine-tuning and distillation method. Therefore, its specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0057] In some embodiments, the apparatus further includes: a data preprocessing module, configured to collect video data through downhole monitoring equipment and preprocess it, extract candidate data containing key task objectives and label them to form a finely labeled dataset, form an unlabeled dataset from the remaining unlabeled data, and filter publicly available general data adapted to the downhole scenario to form a publicly available general dataset.

[0058] In other embodiments, the apparatus further includes a quality assessment module, used to introduce an assessment model to score and verify the updated distillation dataset, wherein the scoring and verification covers dimensions of labeling accuracy, sample fit and content completeness, and filters out problematic data with labeling errors or substandard quality.

[0059] In one specific embodiment, the device can be deployed in a data center server on the surface of a coal mine or in an edge computing device underground, and realizes a complete process of data acquisition, model training, distillation optimization and enhancement fine-tuning through modular design.

[0060] Specifically, the modules in the device can be implemented by a combination of software, hardware, or firmware. For example, modules 100 to 400 can be implemented by software or firmware stored in memory and executed by a suitable instruction execution system; they can also be implemented in hardware, such as by any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0061] It should be noted that the coal mine underground model fine-tuning and distillation device in this embodiment effectively solves the problems of poor model adaptability and low data utilization in underground coal mine scenarios by constructing a three-level data system of "precisely labeled data + unlabeled data + publicly available general data" and combining supervised fine-tuning, distillation iteration, and enhanced fine-tuning mechanisms. This significantly improves the model's detection accuracy and inference efficiency in complex environments. Simultaneously, through lightweight distillation and iterative optimization mechanisms, the data labeling cost and the deployment threshold for edge devices are reduced, achieving synergistic optimization of high accuracy and low resource consumption.

[0062] A specific embodiment of the present invention also provides a coal mine underground model fine-tuning and distillation device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the above-mentioned coal mine underground model fine-tuning and distillation method.

[0063] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0064] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0065] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0066] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0067] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for fine-tuning and distillation of an underground coal mine model, characterized in that, include: Acquire a finely labeled dataset, an unlabeled dataset, and a publicly available dataset for underground coal mine scenarios. Use the finely labeled dataset to perform initial supervised fine-tuning on the teacher model and the target model to obtain the initial teacher model and the initial target model. The initial teacher model and the initial target model are used to perform multi-channel parallel inference on the unlabeled dataset to generate candidate labeling results. The candidate labeling results are then subjected to consistency verification and quality screening to construct a high-quality distillation dataset. The multi-channel parallel inference generates multiple candidate labeling results with diversity in content and inference path by configuring generation parameters, inference hints, inference structure or context information in a differentiated manner. High-quality pseudo-labels are screened through a dual verification mechanism including consistency check and underground scene logic matching degree compliance check. The precise labeled dataset, the publicly available general dataset, and the high-quality distillation dataset are integrated to form a comprehensive training dataset. The initial target model is then fine-tuned using a multi-loss fusion method with the comprehensive training dataset to obtain an optimized target model. The multi-loss fusion method comprehensively considers the supervision loss corresponding to the precise labeled data, the distillation loss corresponding to the distillation data, and the consistency loss of the model prediction to collaboratively optimize the model parameters. Based on the optimized target model, the high-quality distillation dataset is iteratively updated and its quality is evaluated. The updated data is then used to construct an enhanced fine-tuning dataset containing logical guidance information. The optimized target model is then enhanced and fine-tuned to obtain the final model adapted to underground coal mine edge equipment.

2. The method according to claim 1, characterized in that, The process of acquiring a precisely labeled dataset, an unlabeled dataset, and a publicly available dataset representing a coal mine underground scenario, and then using the precisely labeled dataset to perform initial supervised fine-tuning of the teacher model and the target model to obtain the initial teacher model and the initial target model, includes: Video data is collected and preprocessed by downhole monitoring equipment. Candidate data containing key task objectives are extracted and labeled to form a finely labeled dataset. The remaining unlabeled data is used to form an unlabeled dataset. Publicly available general data that is adapted to the downhole scenario is selected to form a publicly available general dataset. A model whose feature extraction capability meets the accuracy requirements is selected as the teacher model, and a lightweight model is selected as the target model. The teacher model and the target model are simultaneously subjected to initial supervised fine-tuning based on the precision dataset, and some low-level parameters are frozen during the training process. Set training hyperparameters, use a loss function to calculate the deviation between the model prediction results and the precise label to guide the model parameter update, and save the model parameters with basic downhole task processing capabilities through iterative evaluation to obtain the initial teacher model and the initial target model.

3. The method according to claim 2, characterized in that, The step of using a loss function to calculate the deviation between the model prediction results and the precise labels to guide model parameter updates includes: In each training iteration, the sample labels and the predicted probability distribution of the model output are obtained for the current batch of samples. The sample labels and the predicted probability distribution are substituted into the loss function to perform the calculation, and the supervision loss value that represents the deviation between the model prediction result and the label is calculated. Based on the supervised loss value, the backpropagation algorithm is executed to calculate the gradient information of the parameters of each layer of the model, and the gradient information is used to update the network weight parameters of the unfrozen parts of the teacher model and the target model; The updated network weight parameters are used in the next training iteration until the preset number of iterations or model convergence conditions are reached.

4. The method according to claim 1, characterized in that, The process involves using the initial teacher model and initial target model to perform multi-channel parallel inference on the unlabeled dataset to generate candidate labeling results, and then performing consistency verification and quality screening on the candidate labeling results to construct a high-quality distillation dataset. Specifically, the multi-channel parallel inference generates multiple candidate labeling results with diversity in content and inference paths by configuring generation parameters, inference hints, inference structures, or contextual information differently. High-quality pseudo-labels are screened through a dual verification mechanism including consistency checks and compliance checks for the logical matching degree of the downhole scenario. The same unlabeled input data from the unlabeled dataset is simultaneously input into an inference system configured with multiple parallel generation channels. Each parallel generation channel performs independent inference on the unlabeled input data through differentiated configurations of generation parameters, inference hints, inference structure, or context information, generating multiple candidate labeling results with diversity in content and inference path. The multiple candidate annotation results are input into the quality assurance module, and a dual verification mechanism based on consistency constraints is executed to eliminate low-quality annotations that do not meet the quality standards. The dual verification mechanism includes a consistency check by comparing the multiple candidate annotation results with each other and a compliance check on the matching degree between the annotation results and the downhole scene logic. Candidate annotations that pass the dual verification mechanism are selected as high-quality pseudo-labels. These high-quality pseudo-labels are then associated with and stored with the unlabeled input data to construct a high-quality distillation dataset.

5. The method according to claim 1, characterized in that, The integrated training dataset is formed by fusing the precision-calibrated dataset, the publicly available dataset, and the high-quality distillation dataset. This integrated training dataset is then used to perform secondary supervised fine-tuning of the initial target model using multi-loss fusion, resulting in an optimized target model. The multi-loss fusion comprehensively considers the supervision loss corresponding to the precision-calibrated data, the distillation loss corresponding to the distillation data, and the consistency loss of model predictions to collaboratively optimize model parameters, including: The precision dataset, the publicly available dataset, and the high-quality distillation dataset are mixed and spliced ​​together according to a preset ratio to form a comprehensive training dataset containing multi-source heterogeneous data. Remove some of the frozen parameters from the initial target model, and perform secondary supervised fine-tuning training on the unfrozen initial target model using the comprehensive training dataset; During training, a fusion loss function is applied to comprehensively consider the supervision loss corresponding to the labeled data, the distillation loss corresponding to the distillation data, and the consistency loss of the model prediction, so as to optimize the model training effect. The model performance is evaluated through periodic iterations. When the model performance reaches the preset standard and tends to stabilize after several consecutive rounds, the secondary supervised fine-tuning is stopped, and the optimized target model parameters are saved as the optimized target model.

6. The method according to claim 5, characterized in that, The application of a fusion loss function during training, which comprehensively considers the supervision loss corresponding to the labeled data, the distillation loss corresponding to the distillation data, and the consistency loss of the model prediction, to optimize the model training effect includes: In each training iteration, the supervision loss generated by the finely labeled data, the distillation loss generated by the distillation data, and the consistency loss between the model prediction results are calculated separately. Obtain the preset loss weight coefficients, multiply the supervision loss, the distillation loss and the consistency loss by their respective loss weight coefficients, perform an addition operation, and substitute them into the fusion loss function to calculate the total loss value; Based on the total loss value, perform gradient backpropagation to calculate the gradient update amount of each trainable parameter in the initial target model, and use the gradient update amount to adjust the model network weights to minimize the total loss value. The adjusted model network weights are used in the next iteration of training until the preset stopping condition is met.

7. The method according to claim 1, characterized in that, The process involves iteratively updating and evaluating the quality of the high-quality distillation dataset based on the optimized target model, constructing a reinforced fine-tuning dataset containing logical guidance information using the updated data, and then further refining the optimized target model to obtain the final model adapted to underground coal mine edge equipment. The optimized target model is used to re-perform inference labeling on the data in the high-quality distillation dataset to update the pseudo-labels; An evaluation model is introduced to score and verify the updated distillation dataset. The scoring and verification covers the dimensions of labeling accuracy, sample fit and content completeness, and filters out problematic data with labeling errors or substandard quality. The problematic data is re-input into the multi-channel parallel inference process for parallel inference and quality verification again, forming an iterative closed loop until the proportion of problematic data drops below a preset threshold or reaches a preset number of iterations. Based on the needs of underground coal mine monitoring tasks, a thought chain reasoning template adapted to the underground scenario is constructed. Samples from the precision-standard dataset and the iteratively updated distillation dataset are selected and added to the thought chain reasoning template to construct an enhanced fine-tuning dataset containing logical guidance information. The target model after iterative optimization is subjected to reinforcement fine-tuning using the aforementioned reinforcement fine-tuning dataset to obtain the final model adapted to underground edge equipment in coal mines.

8. A fine-tuning and distillation device for an underground coal mine model, characterized in that, include: The initial fine-tuning module is used to acquire the precise labeled dataset, the unlabeled dataset, and the publicly available general dataset for the underground coal mine scenario. The precise labeled dataset is used to perform initial supervised fine-tuning on the teacher model and the target model to obtain the initial teacher model and the initial target model. The distillation data construction module is used to perform multi-channel parallel inference on the unlabeled dataset using the initial teacher model and the initial target model to generate candidate labeling results, and to perform consistency verification and quality screening on the candidate labeling results to construct a high-quality distillation dataset. The multi-channel parallel inference generates multiple candidate labeling results with diversity in content and inference path by configuring generation parameters, inference hints, inference structure or context information in a differentiated manner, and filters high-quality pseudo-labels through a dual verification mechanism including consistency check and underground scene logic matching compliance check. The secondary fine-tuning module is used to fuse the precision-calibrated dataset, the publicly available general dataset, and the high-quality distillation dataset to form a comprehensive training dataset. The initial target model is then subjected to secondary supervised fine-tuning with multi-loss fusion using the comprehensive training dataset to obtain the optimized target model. The multi-loss fusion comprehensively considers the supervision loss corresponding to the precision-calibrated data, the distillation loss corresponding to the distillation data, and the consistency loss of the model prediction to collaboratively optimize the model parameters. The enhancement and fine-tuning module is used to iteratively update and evaluate the quality of the high-quality distillation dataset based on the optimized target model, and to construct an enhancement and fine-tuning dataset containing logical guidance information using the updated data, thereby enhancing and fine-tuning the optimized target model to obtain the final model adapted to underground coal mine edge equipment.

9. The apparatus according to claim 8, characterized in that, The distillation data construction module includes: The annotation candidate unit is used to simultaneously input the same unlabeled input data from the unlabeled dataset into an inference system configured with multiple parallel generation channels. Each of the parallel generation channels performs independent inference on the unlabeled input data through differentiated configurations of generation parameters, inference hints, inference structures, or contextual information, generating multiple candidate annotation results with diversity in content and inference paths. The annotation filtering unit is used to input the multiple candidate annotation results into the quality assurance module and execute a dual verification mechanism based on consistency constraints to eliminate low-quality annotations that do not meet the quality standards. The dual verification mechanism includes a consistency check by comparing the multiple candidate annotation results with each other and a compliance check by checking the matching degree between the annotation results and the downhole scene logic. A high-quality distillation dataset construction unit is used to filter candidate annotation results that pass the dual verification mechanism as high-quality pseudo-labels, associate and store the high-quality pseudo-labels with the unlabeled input data, and construct a high-quality distillation dataset.

10. A coal mine underground model fine-tuning and distillation device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the coal mine underground model fine-tuning and distillation method as described in any one of claims 1 to 7.

Images