Intelligent mine multi-subsystem cooperative comprehensive management method based on federal transfer learning

Abstract

The invention discloses an intelligent mine multi-subsystem cooperative comprehensive management method based on federal transfer learning, and the method comprises the steps: carrying out the pre-training of a deep neural network through an industrial equipment data set and historical desensitization data based on a central server, and obtaining an initial global feature extraction model, distributing global model parameters of the initial global feature extraction model to a plurality of local nodes; receiving global model parameters based on each local node, constructing a localized model, and performing local adaptation training on the localized model by using local private data; feature extraction layer parameters obtained after adaptive training of the localized model are uploaded to a central server, the central server aggregates the feature extraction layer parameters uploaded by all the local nodes, updates the current global feature extraction model, generates a new global feature extraction model and distributes the new global feature extraction model to each local node, and the local nodes send the new global feature extraction model to the local nodes. And repeating the process until the model converges to obtain a collaborative comprehensive management model of the collaborative comprehensive management system based on the global feature extraction model after evolution.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent control technology, and in particular to a collaborative management method for multi-subsystem intelligent mines based on federated transfer learning. Background Technology

[0002] The intelligent mine integrated management and control platform is a product of the deep integration of new-generation information technologies such as the Industrial Internet, big data, and artificial intelligence with mining operations. However, most mining enterprises are still in the early stages of intelligentization, with their main achievements limited to a certain degree of automation and informatization. Various production subsystems, such as personnel positioning, ventilation, drainage, and equipment monitoring, were mostly built independently by different suppliers at different times. While they achieved preliminary data collection and localized monitoring, at the platform level, they primarily functioned as centralized data display and simple alarms, essentially remaining "data dashboards" lacking in-depth analysis and intelligent decision-making capabilities.

[0003] This construction model has led to a widespread phenomenon of "data silos." Due to the different technical standards and data protocols used by various subsystems, data cannot be effectively shared and exchanged. Data from systems such as ventilation, drainage, and production are isolated, making correlation analysis difficult and hindering the formation of a holistic understanding. While some leading companies have begun to introduce AI technology, its applications are relatively singular and isolated, such as using computer vision for belt misalignment detection or using sensor data for equipment fault diagnosis. These applications are typically "siloed," developed for specific problems, and fail to integrate with the entire production process and other systems, resulting in limited levels of intelligence. Meanwhile, data security and privacy concerns are increasingly prominent. Mine production data involves core corporate secrets, and subsystem data also relates to supplier trade secrets. Traditional centralized data processing methods require all data to be uploaded to a central server, causing mining companies and equipment suppliers to be cautious about data sharing for security reasons. This significantly restricts data integration and value extraction.

[0004] Moving from "integration" to "fusion" is the core development direction for the future of intelligent mine integrated management and control platforms. The platform will no longer be limited to the simple accumulation and display of data, but will instead, based on an industrial internet architecture, promote the deep integration of data, models, algorithms, and applications. By building a unified data platform, it will break down system barriers and achieve cross-domain data collaboration and intelligent support. Artificial intelligence applications will also expand from single-point scenarios to full-chain global optimization covering production, operation, and energy efficiency, achieving multi-system collaboration and intelligent scheduling across "mining-excavation-transportation-communication." Digital twin technology will use high-fidelity 3D models to achieve real-time mapping and interaction between the physical mine and the virtual world, supporting simulation, fault backtracking, and emergency drills.

[0005] Despite a promising future, the current intelligent transformation of mines still faces multiple challenges. The problem of information silos is particularly prominent, becoming a core bottleneck restricting development and hindering comprehensive production scheduling, safety management, and emergency command, severely limiting management efficiency. The intelligent decision-making level of systems is generally low, with most only possessing monitoring and alarm functions, lacking predictive, early warning, and pre-control capabilities based on multi-source fusion data. This leads to delayed identification of safety hazards, and equipment maintenance still relies on periodic inspections or post-incident handling, failing to achieve state-based predictive maintenance, resulting in unplanned downtime and production losses. In emergency scenarios, the collaborative linkage capabilities between systems are significantly insufficient, failing to automatically trigger related systems to perform operations such as ventilation adjustment, power outages, and evacuation, still relying on manual intervention and missing crucial response opportunities. Furthermore, traditional centralized architectures require significant investment in central servers and bandwidth resources to transmit, store, and process massive amounts of data, resulting in high construction and maintenance costs. At the market demand level, the construction of smart mines is strongly driven by multiple factors, including policy, safety, economics, and the value of data. Enterprises urgently need new comprehensive management and control technology solutions that can address the aforementioned core issues. Summary of the Invention

[0006] The purpose of this invention is to provide a collaborative management method for multiple subsystems in smart mines based on federated transfer learning. By introducing a collaborative training mechanism of "centralized pre-training - localized customization - hierarchical federated collaboration", it effectively breaks down data silos while ensuring that the data of each mine subsystem does not leave the domain, realizes cross-system knowledge fusion and collaborative intelligent decision-making, and significantly reduces system communication overhead.

[0007] To address the aforementioned technical problems, a first aspect of this invention provides a collaborative management method for multiple subsystems in a smart mine based on federated transfer learning. The collaborative management system includes a central server and local nodes corresponding one-to-one with several mine subsystems. The method includes the following steps: Based on the central server, a deep neural network is pre-trained using industrial equipment datasets and historical de-identified data to obtain an initial global feature extraction model, and the global model parameters of the initial global feature extraction model are distributed to several local nodes. Based on the global model parameters received by each local node, a localized model is constructed, and the local private data of the local node is used to perform local adaptation training on the corresponding localized model. Based on the central server and all the local nodes, collaborative training is performed through a federated learning framework. Each local node uploads the feature extraction layer parameters obtained after the localized model adaptation training to the central server. The central server aggregates the feature extraction layer parameters uploaded by all the local nodes, updates the current global feature extraction model, generates a new global feature extraction model, and distributes it to each local node. This process is repeated until the model converges, resulting in the collaborative management system's collaborative management model based on the evolved global feature extraction model.

[0008] Further, the collaborative training based on the central server and all local nodes, using a federated learning framework, involves each local node uploading the feature extraction layer parameters obtained after the localized model adaptation training to the central server. The central server aggregates the feature extraction layer parameters uploaded by all local nodes, updates the current global feature extraction model, generates a new global feature extraction model, and distributes it to each local node. This process is repeated until the model converges, including: Based on each local node, the corresponding localized model is trained and updated using the local private data, and the updated feature extraction layer parameters of the localized model are uploaded to the central server. Based on the feature extraction layer parameters received by the central server from all the local nodes, the federated averaging algorithm is used to aggregate them to obtain a global feature extraction model, and the global model parameters of the global feature extraction model are distributed to all the local nodes. The feature extraction layer parameters of the corresponding localized model of each local node are replaced with the global model parameters of each local node, while the local input adaptation layer and task-specific output layer parameters of the localized model are retained. Repeat the above process until the global feature extraction model converges.

[0009] Further, the step of aggregating the feature extraction layer parameters uploaded by all local nodes received by the central server using a federated averaging algorithm to obtain a global feature extraction model includes: Based on the feature extraction layer parameters of encrypted transmission received by each local node from the central server, and the integrity verification and decryption processing of the feature extraction layer parameters, the plaintext feature extraction layer parameters of each local node are obtained. Obtain the data volume factor, data quality factor, and model contribution factor of the local private data used for training by each local node in this training cycle, and calculate the plaintext feature extraction layer parameters of the corresponding local node and assign the corresponding aggregation weights. Normalize all the aggregate weights to obtain the normalized weight value of each local node, and calculate the weighted average of the plaintext feature extraction layer parameters of all the local nodes to obtain the global model parameters of the global feature extraction model. The global feature extraction model is constructed based on the global model parameters.

[0010] Further, obtaining the data volume factor of the local private data used for training by each local node in this training cycle includes: Based on the amount of training data used by each local node in this training cycle, the original data volume factor of the local node is obtained, where the amount of training data refers to the number of data samples that the local node actually participates in model training in this training. The original data volume factors of all local nodes are normalized to adjust the numerical range of each original data volume factor to a consistent scale, thus obtaining the data volume factor of each local node.

[0011] Further, obtaining the data quality factor of the local private data used for training by each local node in this training cycle includes: The local private data of the local node is input into the feature extraction layer of the corresponding localization model to obtain the corresponding feature vector; Calculate the information metric of the feature vector, which includes average mutual information or feature sparsity, wherein average mutual information is used to quantify the statistical dependence between feature vectors, and feature sparsity is used to quantify the distribution of non-zero elements in the feature vector. The ratio of outliers in the local private data is evaluated, and the outliers are identified by the isolated forest algorithm or an outlier detection method based on the confidence level of the local model prediction. Based on the information metric and the outlier ratio, a data quality factor is obtained through weighted calculation. The information metric is positively correlated with the data quality factor, and the outlier ratio is negatively correlated with the data quality factor.

[0012] Further, obtaining the model contribution factor of the local node includes: The historical contribution data of the local node in the past multiple rounds of aggregation is obtained from the contribution decay memory queue. The historical contribution data includes the instantaneous improvement of the performance index of the global model on the standard validation set after the feature extraction layer parameters of the local node are aggregated into the global model. Based on the historical contribution data, the model contribution factor of the local node is calculated using the exponential weighted average method, where recent historical contributions are given higher weights. The model contribution factors of all local nodes are normalized to adjust the numerical range of the model contribution factors to a consistent scale.

[0013] Further, the calculation of the plaintext feature extraction layer parameters of the corresponding local node and the allocation of corresponding aggregation weights includes: Based on the data volume factor, data quality factor, and model contribution factor, the initial aggregation weight of each local node is calculated using the weighted geometric average method. Based on the safety risk level of the mining subsystem corresponding to the local node in the current operating cycle, the initial aggregation weight is dynamically adjusted, and the safety risk level is positively correlated with the initial aggregation weight. The aggregate weights of all the dynamically adjusted local nodes are normalized to obtain the aggregate weight values ​​used for calculating the weighted geometric mean.

[0014] Further, uploading the updated feature extraction layer parameters of the localized model to the central server includes: The feature extraction layer parameters are then homomorphically encrypted or differentially private before being uploaded.

[0015] Further, the step of receiving the global model parameters from each local node, constructing a localized model, and using the local private data of the local node to perform local adaptation training on the corresponding localized model includes: Based on the global model parameters of the initial global feature extraction model, a local base model with the same network structure is constructed on the local node. Based on the local base model, an input adaptation layer for adapting to local data dimensions and a task-specific output layer configured according to the local management task type are added to obtain the localized model. While keeping the model parameters of the local base model frozen, the local private data of the local node is used to train the input adaptation layer and the task-specific output layer in the localized model until the loss function of the newly added network layer reaches a stable state, thus completing the local adaptation training.

[0016] Furthermore, the step of training the input adaptation layer and task-specific output layer in the localized model using the local node's local private data while keeping the model parameters of the local base model frozen, until the loss function of the newly added network layer reaches a stable state, includes: A forward propagation computation process is constructed on the local node, and the local private data is input into the localized model. The model then passes through the local base model in a parameter-frozen state, the input adaptation layer in a trainable state, and the task-specific output layer in sequence to obtain the model prediction output. Calculate the difference between the model's predicted output and the true label, and calculate the loss value for the current training state based on this difference using a loss function; The backpropagation algorithm is executed, and the gradient update of the backpropagation algorithm is limited to the model parameters of the input adaptation layer and the task-specific output layer; Based on the gradient calculated by the backpropagation algorithm, the model parameters of the input adaptation layer and the task-specific output layer are updated using the stochastic gradient descent algorithm. Repeat the above training process until the change in the loss function value is less than the preset convergence threshold in multiple consecutive training cycles, thus completing the training of the new network layer.

[0017] Furthermore, the task-specific output layer configured according to the local management task type includes: Linear output layer for regression tasks; A Softmax output layer used for classification tasks; or, An autoencoder output structure for anomaly detection tasks.

[0018] Furthermore, the initial global feature extraction model, based on a central server and pre-trained using industrial equipment datasets and historical de-identified data, includes: The central server collects industrial equipment operation data and historical mine monitoring data, and de-identifies the historical mine monitoring data to form a pre-training dataset. A deep neural network structure is constructed, the deep neural network is trained using the pre-training dataset, the predicted output is calculated through forward propagation, the loss function is calculated based on the difference between the predicted output and the true label, and the weight parameters of the deep neural network are updated through the backpropagation algorithm. The training process is repeated until the change in the loss function value of the deep neural network is lower than a preset threshold in consecutive preset rounds. The trained deep neural network is then used as the initial global feature extraction model.

[0019] Furthermore, the construction of the deep neural network structure, training the deep neural network using the pre-training dataset, calculating the predicted output through forward propagation, calculating the loss function based on the difference between the predicted output and the true label, and updating the weight parameters of the deep neural network through the backpropagation algorithm, includes: Construct a deep neural network structure containing multiple feature extractors, each of which is composed of a combination of a linear transformation unit and a nonlinear activation function; The sample data in the pre-training dataset is input into the deep neural network. Through the forward propagation process of performing linear transformation and nonlinear activation layer by layer, the predicted output of the deep neural network corresponding to the input sample data is obtained. Calculate the difference measure between the predicted output and the true label of the sample, and calculate the loss value in the current training state based on this difference measure and a preset loss function; The backpropagation algorithm is executed based on the loss value to calculate the gradient value of the loss value with respect to the weight parameters of each layer of the deep neural network. Based on the calculated gradient value, all weight parameters of the deep neural network are updated using the stochastic gradient descent algorithm.

[0020] Furthermore, the mining subsystem is an independently operating and data-heterogeneous subsystem within a smart mine, including at least one of personnel positioning, ventilation, drainage, and equipment monitoring.

[0021] Accordingly, a second aspect of the present invention provides an electronic device, including: at least one processor; and a memory connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to execute the above-described intelligent mine multi-subsystem collaborative management method based on federated transfer learning.

[0022] Accordingly, a third aspect of the present invention provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the above-described intelligent mine multi-subsystem collaborative management method based on federated transfer learning.

[0023] The above-described technical solutions of the embodiments of the present invention have the following beneficial technical effects: 1. By using a federated learning framework, each mine subsystem can effectively integrate knowledge from isolated subsystems such as personnel positioning, ventilation, and drainage without uploading local private raw data. This is achieved by exchanging and aggregating encrypted intermediate layer feature vectors, thus training an enhanced model with a global perspective. This breaks the inherent "data silo" dilemma. When optimizing the ventilation system, personnel positioning information can be comprehensively considered. When diagnosing equipment faults, the operating status of the drainage system can be correlated. Ultimately, this significantly improves the accuracy and globality of intelligent decision-making tasks such as safety risk warning and production process optimization. 2. The system adopts a mechanism of "centralized pre-training combined with hierarchical federated collaboration". During the federated iteration process, each local node only needs to upload the feature extraction layer parameters of the model to the central server, while the massive amount of raw data and the local private input / output layer parameters are kept locally. This extreme compression of the amount of uploaded data reduces the total communication load of the system from the terabit level of traditional centralized learning or full-parameter federated learning to the gigabit level, which greatly alleviates the bandwidth pressure on the mining industrial ring network and enables the system to access and expand more subsystems at a lower cost. 3. Through the model architecture design of "shared feature layer + private task layer", supplemented by pre-training initialization and phased training strategies, the final collaborative model has two advantages: on the one hand, the shared feature extraction layer obtained by federated aggregation optimization integrates the common knowledge of all subsystems and has a powerful and universal feature representation capability; on the other hand, the local input adaptation layer and task-specific output layer retained by each subsystem enable it to accurately adapt to its own unique data structure and management task, thereby ensuring the high performance of the global model while perfectly meeting the personalized needs of different subsystems and achieving precise management of "harmony in diversity". Attached Figure Description

[0024] Figure 1 This is a flowchart of the collaborative management method for multiple subsystems in a smart mine based on federated transfer learning provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the collaborative management principle of a multi-subsystem smart mine based on federated transfer learning, provided in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0026] Please refer to Figure 1 and Figure 2 The first aspect of this invention provides a collaborative management method for multiple subsystems in a smart mine based on federated transfer learning. The collaborative management system includes a central server and local nodes corresponding one-to-one with several mine subsystems. The method includes the following steps: Step S100: Based on the central server, the deep neural network is pre-trained using industrial equipment dataset and historical de-identified data to obtain an initial global feature extraction model, and the global model parameters of the global feature extraction model are distributed to several local nodes.

[0027] On the central server side, centralized pre-training of the deep neural network is first performed. This process utilizes a well-organized dataset from a standard industrial equipment library, along with de-identified and generalized historical monitoring data from the mine, as training samples. This data is ensured to be free of any identifiable information specific to any particular mine. The pre-training task aims to build a global feature extraction model with strong generalization capabilities. This model typically employs a deep convolutional network or transformer architecture, learning on massive amounts of industrial data to enable its lower network layers to capture the general intrinsic patterns and feature representations of multi-source signals such as equipment vibration, temperature, and pressure. For example, it can extract common spectral features or temporal patterns related to equipment health status and workload. After training, the central server distributes all parameters (i.e., weights and bias matrices) of the generated initial global feature extraction model to the local nodes of each mine subsystem through a secure network channel within the mine. The parameters distributed here form the unified initialization basis for subsequent federated learning, ensuring that all nodes start with a high level of feature extraction capability, rather than being randomly initialized from scratch, thus laying a solid foundation for subsequent collaborative training.

[0028] Step S200: Based on the global model parameters received by each local node, a localized model is constructed, and the localized model is trained locally using the local private data of the local node.

[0029] After receiving the pre-trained model parameters, each local node first instantiates a local base model with the same network structure. Then, it customizes the model according to the specific needs of its subsystem: for data dimension mismatches, such as the number of sensor channels in a ventilation subsystem not matching the input requirements of the pre-trained model, the node adds a fully connected layer as an input adapter to project local data to the desired dimensions. Simultaneously, based on the specific control task, the node constructs and connects dedicated output layers; for example, a linear regression output layer is used in the drainage subsystem to predict water level changes, a softmax classification output layer is used in the personnel positioning subsystem to identify personnel areas, and an autoencoder is used to reconstruct the output structure in the equipment monitoring subsystem to detect anomalies. In the subsequent local adaptation training phase, the local node uses the backpropagation algorithm but strictly freezes the pre-trained feature extraction layer parameters using gradient masking. Supervised training is performed on the newly added input adaptation layer and task output layer using only local private sensor data streams and operation records. This process aims to quickly adjust the model to adapt to the local data distribution and task objectives while preserving the general knowledge contained in the pre-trained model.

[0030] Step S300: Based on the central server and all local nodes, collaborative training is performed through a federated learning framework. Each local node uploads the feature extraction layer parameters obtained after local model adaptation training to the central server. The central server aggregates the feature extraction layer parameters uploaded by all local nodes, updates the current global feature extraction model, generates a new global feature extraction model, and distributes it to each local node. This process is repeated until the model converges, resulting in a collaborative management system model based on the evolved global feature extraction model.

[0031] Then, the hierarchical federated collaborative training phase begins, which is a multi-round iterative process. In each iteration, each local node first unfreezes the feature extraction layer in its localized model and uses the latest local private data batch to train the entire model end-to-end, updating all parameters, including those of the feature extraction layer. After training, the local node does not upload all model parameters, but selectively encrypts the updated feature extraction layer parameters and uploads them to the central server, while keeping the local input adaptation layer and task-specific output layer parameters as private parameters locally. After collecting the encrypted parameters uploaded by all nodes, the central server decrypts and securely aggregates them, using a federated averaging algorithm to assign aggregation weights to each node based on the amount of data it participated in during training, calculating the next-generation global feature extraction model parameters. The server then distributes these updated global feature extraction model parameters to all nodes, and each local node immediately replaces its local feature extraction layer parameters with them, while retaining its personalized local layer parameters, thus completing one round of knowledge fusion. This process is repeated until the performance evaluation metrics of the global feature extraction model stabilize on the validation set, ultimately enabling each mining subsystem to obtain a robust collaborative model that shares global optimized feature extraction capabilities while also possessing local task specificity.

[0032] In the hierarchical federated collaborative training phase, the localized model built on local nodes adopts a modular hierarchical architecture, which consists of three core parts: a shared global feature extraction layer, a private local input adaptation layer, and a task-specific output layer. The global feature extraction layer forms the basic backbone network of the model, responsible for learning generalized feature representations from raw data. Its parameters are uniformly distributed and updated by the central server, achieving knowledge sharing across subsystems. The input adaptation layer, as a pre-processing module, is responsible for mapping the unique, multi-dimensional local private data of each subsystem to the standardized input space required by the feature extraction layer. The task-specific output layer, as a post-processing module, is customized according to the specific management task type of the subsystem. This hierarchical design enables efficient parameter separation during federated training: only the parameters of the feature extraction layer participate in the aggregation and updating of the central server, while the parameters of the input adaptation layer and the task-specific output layer remain locally, ensuring both data privacy and task specificity of each subsystem, and achieving the collaborative evolution of global knowledge.

[0033] Through the above steps, an efficient, safe, and intelligent collaborative management and control system was built in the complex environment of the mine. It effectively broke down the information barriers between various heterogeneous subsystems, realized cross-domain knowledge sharing and integration under the protection of data privacy, significantly improved the system's comprehensive decision-making capabilities and automation level in areas such as safety early warning, production optimization, and equipment maintenance, and at the same time greatly reduced the communication and computing resource consumption caused by traditional centralized processing.

[0034] Specifically, in step S100, based on the central server, the deep neural network is pre-trained using industrial equipment datasets and historical de-identified data to obtain the global feature extraction model of the initial global model, including: Step S110: Collect industrial equipment operation data and historical mine monitoring data based on the central server, and perform anonymization processing on the historical mine monitoring data to form a pre-training dataset.

[0035] On the central server side, the first step is to construct a representative and comprehensive pre-trained dataset. The data sources mainly consist of two parts: first, standardized equipment operation datasets obtained from standard industrial equipment manufacturers. These datasets cover equipment condition monitoring data under various typical operating conditions, including motor vibration spectra, bearing temperature curves, and hydraulic system pressure waveforms; second, anonymized data extracted from historical monitoring databases of multiple typical mines. This data originally included ventilation speed records, drainage pump station operation logs, and conveyor belt monitoring video frames. For historical mine data, a strict anonymization process is required, including removing all direct identifiers (such as mine numbers and specific coordinates), obfuscating timestamps, normalizing numerical monitoring data to standard intervals, and using data generalization techniques to replace specific equipment models with equipment type classifications. This ensures that the processed data cannot be traced back to a specific mine or equipment, while preserving the statistical characteristics and physical patterns of the industrial monitoring data.

[0036] The pre-training dataset, generated through the aforementioned processing, encompasses both the general operational characteristics of industrial equipment and data patterns specific to the mining environment, while completely removing sensitive information related to specific mine entities. The dataset is typically organized using time-series segments or equipment status samples, with each sample bearing a corresponding operating condition label or equipment status label, providing a high-quality training foundation for subsequent supervised learning of deep neural networks. This data preparation method satisfies the initial model training data volume requirements while fundamentally avoiding the privacy risks that might arise from directly using sensitive operational data from various mines during the pre-training phase.

[0037] Step S120: Construct a deep neural network structure, train the deep neural network using a pre-trained dataset, calculate the predicted output through forward propagation, calculate the loss function based on the difference between the predicted output and the true label, and update the weight parameters of the deep neural network through the backpropagation algorithm.

[0038] The constructed deep neural networks typically employ a hierarchical feature extraction architecture, such as a deep convolutional network consisting of multiple convolutional layers, pooling layers, and fully connected layers, or a transformer-encoder stack with a self-attention mechanism. The core design goal of the network is to build a model capable of extracting abstract feature representations layer by layer from raw industrial sensor data or preprocessed monitoring images. The bottom layers learn to detect local features such as edges and periodic patterns; the middle layers learn to combine these local features to form equipment state patterns; and the higher layers learn to extract global feature representations highly correlated with equipment health status and operating efficiency. During training, the network receives samples from a pre-training dataset, calculates the activation values ​​of each layer through forward propagation, and finally obtains the predicted output. This predicted output is compared with the true label of the sample, and an appropriate loss function, such as mean squared error or cross-entropy, is used to quantify the degree of difference between the predicted and true values.

[0039] Based on the calculated loss value, the training process employs the backpropagation algorithm to calculate the gradient of the loss function with respect to the weight parameters of each layer of the network, explicitly indicating the direction and magnitude of parameter adjustments. Subsequently, stochastic gradient descent or its variants, such as Adam or RMSprop, are used to update all weight parameters of the network according to the learning rate hyperparameter settings. The entire training process is performed on a GPU cluster using a mini-batch training method. Each time, a batch of samples is randomly selected from the pre-training dataset for forward and backward propagation. Through numerous iterations of parameter updates, the network gradually learns the ability to extract effective feature representations from various industrial equipment data and mine monitoring data, ultimately forming a highly generalizable feature extractor.

[0040] Step S130: Repeat the training process until the change in the loss function value of the deep neural network is lower than the preset threshold in consecutive preset rounds, and use the trained deep neural network as the global feature extraction model.

[0041] Model training is an iterative optimization process that requires continuously executing the entire process of forward propagation, loss calculation, backpropagation, and parameter updates. During training, the changes in the loss function value of the deep neural network on a reserved validation set are continuously monitored. This validation set consists of a portion of data reserved from the pre-training dataset that was not used in training, and is used to objectively evaluate the model's generalization ability. The convergence criterion is set as follows: when the decrease in the loss function value over multiple consecutive training epochs (e.g., 20 consecutive epochs) is less than a preset threshold (e.g., 1e). -5 When the convergence condition is met, it can be determined that the model has been sufficiently trained and has reached convergence, at which point the training process is terminated. This convergence judgment mechanism ensures that the model will neither underfit due to insufficient training nor overfit to specific noise in the training data due to overtraining.

[0042] Once the model reaches convergence, the final trained deep neural network is used as the global feature extraction model. The core value of this model lies in its feedforward network's feature extraction layers. These intermediate layers map high-dimensional, complex raw industrial data to a low-dimensional, dense feature vector space. Data points with similar equipment states or operating conditions cluster together in the feature space, while data from different states are separated. At this point, an initial global model with powerful generalized feature extraction capabilities is complete. This model is not specific to any particular mine or subsystem, but it incorporates general feature representation knowledge learned from diverse industrial scenarios, providing a high-quality initialization foundation for subsequent distribution to various mine subsystems for localized customization and federated collaborative learning.

[0043] Further, in step S120, constructing the deep neural network structure involves training the deep neural network using a pre-training dataset, calculating the predicted output through forward propagation, calculating the loss function based on the difference between the predicted output and the true label, and updating the weight parameters of the deep neural network through the backpropagation algorithm, including: Step S121: Construct a deep neural network structure containing multiple feature extractors, each feature extractor consisting of a combination of a linear transformation unit and a nonlinear activation function.

[0044] When constructing the deep neural network structure, a deep feedforward network architecture consisting of multiple fully connected layers is adopted. Each fully connected layer serves as a basic feature extractor, containing a linear transformation unit and a nonlinear activation function. The linear transformation unit linearly combines the input data using a weight matrix and a bias vector, mapping the input features to a new feature space. Subsequently, a nonlinear activation function, such as the ReLU function, is applied to introduce nonlinear transformation capabilities, enabling the network to learn and represent complex nonlinear patterns in the input data. This combination of "linear transformation + nonlinear activation" units is stacked layer by layer to form a deep network structure. The bottom layer network learns basic feature patterns, the middle layer network combines basic features to form composite features, and the high layer network extracts abstract feature representations highly correlated with equipment status and operating conditions, ultimately constructing a deep neural network model with powerful feature extraction capabilities.

[0045] Step S122: Input the sample data in the pre-training dataset into the deep neural network, and obtain the predicted output of the deep neural network corresponding to the input sample data by performing the forward propagation process of linear transformation and nonlinear activation layer by layer.

[0046] Sample data from the pre-training dataset, such as equipment vibration spectra and temperature time-series data, are input into the constructed deep neural network to initiate the forward propagation computation process. The input data first enters the network's first-layer feature extractor. After linear transformation and nonlinear activation processing, the output activation values ​​serve as the input for the next layer, sequentially passing through each layer for feature extraction and transformation. In each layer, linear transformation achieves spatial mapping of features through matrix multiplication, while nonlinear activation functions ensure the network can learn complex decision boundaries. Finally, the output layer produces prediction results for the input samples, such as equipment state classification probabilities or operating condition regression values, completing the full transformation process from raw data to high-level abstract features.

[0047] Step S123: Calculate the difference measure between the predicted output and the true label of the sample, and calculate the loss value in the current training state based on this difference measure and a preset loss function.

[0048] By comparing the differences between the network's predicted output and the true labels of the samples, an appropriate loss function is used to quantify the prediction error of the current model. For classification tasks, the cross-entropy loss function is used to calculate the difference between the predicted probability distribution and the true label distribution; for regression tasks, the mean squared error loss function is used to measure the squared distance between the predicted value and the true value. The loss value is calculated based on the overall error of the batch training samples. The prediction results obtained through forward propagation are compared with the labeled real-world data. The resulting loss value not only reflects the degree of fit of the current model to the training data but also provides clear guidance on the direction and magnitude of subsequent parameter optimization.

[0049] Step S124: Execute the backpropagation algorithm based on the loss value to calculate the gradient value of the loss value with respect to the weight parameters of each layer of the deep neural network, and update all weight parameters of the deep neural network based on the calculated gradient value using the stochastic gradient descent algorithm.

[0050] Based on the calculated loss value, the backpropagation algorithm is executed, and the gradient of the loss function with respect to the weight parameters of each layer of the network is calculated using the chain rule. This process starts from the output layer and propagates the error signal backward layer by layer, accurately calculating the contribution of each weight parameter to the overall loss. After obtaining the gradient information, the parameters are updated using the stochastic gradient descent algorithm. The update step size is controlled according to the preset learning rate hyperparameter, and the weight parameters of each layer are adjusted along the opposite direction of the gradient, gradually reducing the loss function value. This optimization process continuously adjusts the network parameters through multiple iterations, making the model's prediction results gradually approach the true labels, ultimately achieving a continuous improvement in the network's feature extraction capability.

[0051] Specifically, step S200, which involves receiving global model parameters for each local node, constructing a localized model, and using the local private data of the local node to perform local adaptation training on the corresponding localized model, includes: Step S210: Based on the global model parameters of the global feature extraction model, construct a local basic model with the same network structure on the local node.

[0052] After receiving the global feature extraction model parameters from the central server, the local nodes of each mining subsystem first instantiate a local base model in their local computing environment with the exact same network architecture as the pre-trained model on the server. This process involves precisely loading the received model parameters into the corresponding layers of the local model, including the weight matrices and bias vectors of each convolutional layer and the parameters of the batch normalization layer, ensuring that the local base model has the same feature extraction capabilities as the pre-trained model on the central server during the initialization phase. Taking the ventilation subsystem as an example, the local base model established by its local nodes will contain a complete feature extraction layer structure, which already has the ability to extract general features from various industrial sensor data.

[0053] The construction of the local base model not only achieves accurate reproduction of model parameters, but more importantly, establishes a unified feature extraction benchmark. This model inherits the general feature representation capabilities learned by the pre-trained model on diverse industrial data, enabling the mapping of local raw monitoring data to a unified feature space. This initialization method ensures that different subsystems are at the same feature extraction level at the start of federated learning, avoiding inconsistencies in feature representation caused by random initialization and providing a reliable technical foundation for subsequent collaborative training.

[0054] Step S220: Based on the local basic model, add an input adaptation layer for adapting to local data dimensions, and a task-specific output layer configured according to the local management task type to obtain a localized model.

[0055] Building upon the established local base model, each local node customizes its model structure according to the specific needs of its subsystem. First, addressing the potential dimensionality mismatch between local data sources and the input requirements of the pre-trained model, a dedicated input adaptation layer is added to the front end of the local base model. This layer typically employs a fully connected network structure and is responsible for mapping local sensor data in a specific format to the input space desired by the pre-trained model. For example, the drainage subsystem might only have 8 pressure sensor channels, while the pre-trained model requires 16 input channels; the input adaptation layer will achieve this dimensionality transformation through a learnable weight matrix. Simultaneously, based on the specific control tasks of each subsystem, a corresponding task-specific output layer is added behind the feature extraction layer, forming a complete localized model architecture.

[0056] The configuration of task-specific output layers is strictly designed based on the functional requirements of the subsystems. The ventilation subsystem uses a linear regression output layer for airflow prediction, the personnel positioning subsystem uses a Softmax multi-classification output layer for area identification, and the equipment monitoring subsystem uses a reconstructed output layer with an autoencoder structure for anomaly detection. These specially designed output layers, combined with a general feature extraction layer, enable each local node to achieve accurate modeling for specific scenarios while sharing general feature extraction capabilities. This modular architecture maintains the overall system consistency while fully considering the individual needs of each subsystem.

[0057] Step S230: While keeping the model parameters of the local base model frozen, use the local private data of the local node to train the input adaptation layer and the task-specific output layer in the localized model until the loss function of the newly added network layer reaches a stable state, thus completing the local adaptation training.

[0058] During the local adaptation training phase, a staged training strategy is adopted. Parameter freezing technology is used to maintain the non-updateable state of all pre-trained parameters in the local base model, allowing only newly added input adaptation layers and task-specific output layers to participate in the training process. Training utilizes unique local private datasets for each node, including real-time sensor readings and device operating status records. Forward propagation is used to calculate prediction results, and the loss function value is calculated based on the difference between the predicted output and the true label. Backpropagation is then used to update only the parameters of newly added network layers. This training method ensures rapid adaptation to local task requirements while fully preserving the general feature extraction knowledge learned by the pre-trained model.

[0059] During training, the changes in the loss function of newly added network layers are continuously monitored. When the decrease in the loss value over multiple consecutive training cycles is less than a preset threshold, the local adaptation training is considered to have converged. Taking the device monitoring subsystem as an example, its autoencoder output layer learns the normal operating mode of the local device by minimizing the difference between the input data and the reconstructed data. When the reconstruction error stabilizes at a low level, it indicates that the model has fully grasped the normal operating characteristics of the local device. After completing the local adaptation training, each subsystem obtains a personalized model that possesses both general feature extraction capabilities and optimization for specific local tasks.

[0060] Furthermore, the task-specific output layer configured according to the local management task type in step S220 includes: Linear output layers for regression tasks, Softmax output layers for classification tasks, or autoencoder output structures for anomaly detection tasks.

[0061] In the mining subsystem, the regression task primarily targets monitoring scenarios requiring continuous numerical prediction. The linear output layer implements this function through one or more linear transformation units. This output layer receives abstract feature vectors from the feature extraction layer as input, and through linear combination of weight matrices and bias terms, directly outputs one or more continuous predicted values. Taking the ventilation subsystem as an example, this output layer maps the feature vectors output by the general feature extractor to specific airflow prediction values. It optimizes parameters by minimizing the mean square error between the predicted and measured airflow, enabling the model to accurately predict ventilation needs under different operating conditions. In actual deployment, the number of parameters in the linear output layer depends on the dimension of the feature vectors and the number of prediction targets, and is optimized using the standard gradient descent algorithm. During training, the output layer collaborates with the feature extraction layer, but only updates its own parameters, ensuring rapid adaptation to local, specific regression prediction tasks while maintaining general feature extraction capabilities. This design allows the subsystem to accurately complete local continuous value prediction requirements while sharing global knowledge.

[0062] In mine monitoring, classification tasks are widely used for state recognition and pattern classification. The Softmax output layer implements classification through a multi-class logistic regression model. This layer maps the feature vectors output by the feature extraction layer to the dimension of the number of classes through a fully connected layer, and then applies the Softmax function to convert the output values ​​into a probability distribution. Taking the personnel positioning subsystem as an example, this output layer converts the feature vectors into the probability distribution of personnel in various preset areas, and optimizes the parameters through the cross-entropy loss function, enabling the model to accurately identify the area where the personnel are located. During model training, the Softmax output layer and the feature extraction layer together form a complete classification network, but only the parameters of the output layer participate in the update. This design retains the general feature extraction capability of the pre-trained model while quickly adapting to the local classification task requirements. The output layer structure of each subsystem is customized according to the specific number of classification categories to ensure that the model output dimension perfectly matches the actual classification requirements.

[0063] The anomaly detection task requires the model to identify abnormal operating states of equipment or systems. The autoencoder output structure achieves this goal by reconstructing the error. This structure consists of an encoder and a decoder. The encoder further compresses the feature vectors output from the feature extraction layer into a latent representation, while the decoder reconstructs the original feature vectors from the latent representation. Taking an equipment monitoring subsystem as an example, this structure learns the feature patterns of normal equipment states and identifies abnormal states by calculating the reconstruction error. When the reconstruction error exceeds a threshold, an alarm mechanism is triggered. In practical applications, the autoencoder structure is trained using only data under normal operating conditions, learning the feature representation of normal states by minimizing the reconstruction error. This unsupervised learning method is particularly suitable for anomaly detection scenarios because abnormal samples are often difficult to obtain in reality. After training, the structure can effectively distinguish between normal and abnormal states, providing reliable technical support for predictive maintenance of equipment.

[0064] Further, in step S230, while keeping the model parameters of the local base model frozen, the local private data of the local node is used to train the input adaptation layer and the task-specific output layer in the localized model until the loss function of the newly added network layer reaches a stable state, including: Step S231: Construct a forward propagation computation process on the local node, input local private data into the localized model, and sequentially pass through the local base model in the parameter frozen state, the input adaptation layer in the trainable state, and the task-specific output layer to obtain the model prediction output.

[0065] When constructing the forward propagation computation process at the local node, local private data (such as multi-channel pressure sensor readings of the drainage subsystem) is first input into the localized model. The data first passes through the input adaptation layer, which projects the raw data onto a feature space compatible with the pre-trained model using a learnable weight matrix. Subsequently, the data flows into the local base model, which is in a parameter-frozen state. The weights of this model remain read-only during training, performing only feature extraction computation without updating the parameters. Finally, the extracted feature vectors are passed to the task-specific output layer (such as a linear output layer for water level prediction), through which the final model prediction output is obtained. Throughout the forward propagation process, only the parameters of the input adaptation layer and the output layer participate in the computation and retain gradient information. The base model part operates as a fixed feature extractor, ensuring that the knowledge representation obtained through pre-training is not destroyed when adapting to the local task.

[0066] Step S232: Calculate the difference between the model's predicted output and the true label, and calculate the loss value in the current training state based on this difference using the loss function.

[0067] When calculating the difference between the model's predicted output and the true labels, appropriate evaluation methods are used based on the task type. For regression tasks (such as ventilation volume prediction), the mean squared error loss function is used to calculate the average of the squared differences between the predicted and actual monitored values. For classification tasks (such as personnel area identification), the cross-entropy loss function is used to measure the difference between the predicted probability distribution and the true label distribution. The loss value is calculated based on all training samples in the current batch. The prediction results obtained through forward propagation are compared point-by-point with the labeled real data. The resulting comprehensive loss value objectively reflects the degree of fit of the newly added network layer in the current training state, providing an accurate quantitative basis for subsequent parameter optimization.

[0068] Step S233: Execute the backpropagation algorithm. The gradient update of the backpropagation algorithm is limited to the model parameters of the input adaptation layer and the task-specific output layer.

[0069] When executing the backpropagation algorithm, a gradient shielding mechanism strictly limits parameter updates to the input adaptation layer and the task-specific output layer. Starting from the output layer, the algorithm backpropagates the error signal along the computation graph, calculating the partial derivatives of the loss function with respect to the parameters of each layer. When the gradient propagates to the local base model, the gradient flow is truncated to ensure that all parameters in the base model (including convolutional kernel weights, fully connected layer parameters, etc.) do not receive gradient updates. This gradient control mechanism is implemented through the automatic differentiation system of the deep learning framework, setting necessary gradient blocking nodes in the computation graph to precisely control the range of parameter updates while maintaining computational efficiency.

[0070] Step S234: Based on the gradients calculated by the backpropagation algorithm, update the model parameters of the input adaptation layer and the task-specific output layer using the stochastic gradient descent algorithm.

[0071] Based on the gradients calculated through backpropagation, the model parameters of the input adaptation layer and the task-specific output layer are updated using the stochastic gradient descent algorithm. The update process follows the basic principles of gradient descent, shifting the current parameters one step in the opposite direction of the gradient. This step size is controlled by a preset learning rate hyperparameter. Specifically, the weight matrix of the input adaptation layer and the parameters of the output layer are updated according to their corresponding gradient values: the weight parameters are subtracted from the product of the learning rate and the gradient, and the bias parameters are adjusted in the same way. This targeted optimization enables the newly added network layers to quickly learn how to adapt general feature representations to local, specific data distributions and task requirements, while ensuring that the basic feature extraction capability remains unaffected.

[0072] Step S235: Repeat the above training process until the change in the loss function value is less than the preset convergence threshold in multiple consecutive training cycles, and complete the training of the new network layer.

[0073] The training process is repeated, continuously monitoring the changes in the loss function value. When the decrease in the loss value is less than a preset convergence threshold (e.g., 0.001) for multiple consecutive training epochs (e.g., 10 epochs), the newly added network layer is considered to have reached a stable state. The convergence threshold is set according to the accuracy requirements of the specific task. A smaller threshold is set for security early warning tasks requiring high accuracy, while the standard can be appropriately relaxed for general monitoring tasks. After the convergence condition is met, the training process automatically terminates. At this point, the input adaptation layer has learned to effectively map local data to the general feature space, and the task output layer has also completed specialized optimization for the local task, forming a complete model that retains global feature extraction capabilities while adapting to local needs.

[0074] Specifically, in step S300, collaborative training is performed based on the central server and all local nodes using a federated learning framework. Each local node uploads the feature extraction layer parameters obtained after localized model adaptation training to the central server. The central server aggregates the feature extraction layer parameters uploaded by all local nodes, updates the current global feature extraction model, generates a new global feature extraction model, and distributes it to each local node. This process is repeated until the model converges, resulting in a collaborative management system model based on the evolved global feature extraction model, including: Step S310: Train and update the localized model based on local private data for each local node, and upload the updated feature extraction layer parameters of the localized model to the central server.

[0075] Each local node trains a complete localized model based on its local private data. During this phase, the parameters of the local base model are unfrozen, allowing them to participate in training updates along with the input adaptation layer and the task-specific output layer. During training, local nodes use their unique operational data, such as wind pressure sensor data for the ventilation subsystem and water pump operation data for the drainage subsystem, to update all model parameters through a complete forward and backward propagation process. After training, the nodes only extract the parameters of the feature extraction layer from the updated model. These parameters contain feature representation knowledge optimized based on local data and are then uploaded to the central server via an encrypted channel. The parameters of the input adaptation layer and the task-specific output layer remain as local private parameters within the nodes and are not transmitted or shared.

[0076] This selective upload mechanism has multiple advantages: First, it ensures that the core business data and specific configuration parameters of each subsystem do not leave the local environment, effectively protecting trade secrets and operational privacy; second, by transmitting only the feature extraction layer parameters instead of the complete model or raw data, it significantly reduces network transmission load, enabling efficient collaborative learning even under the relatively limited network bandwidth conditions in mines; finally, the feature extraction layer parameters, as intermediate feature representations, have stronger abstraction and privacy protection compared to the raw data, making it difficult to reconstruct the specific business data even without encryption.

[0077] Step S320: Based on the feature extraction layer parameters received by the central server from all local nodes, the federated averaging algorithm is used to aggregate them to obtain a global feature extraction model, and the global model parameters of the global feature extraction model are distributed to all local nodes.

[0078] After receiving the encrypted feature extraction layer parameters from all participating nodes, the central server first decrypts and verifies their integrity to ensure that the parameters have not been tampered with during transmission. Then, a federated averaging algorithm is used to aggregate the parameters. This algorithm assigns a corresponding aggregation weight to each node's parameters based on the amount of data used by each node in this round of training; nodes with larger amounts of data have higher influence in the aggregation process. Specifically, the server first normalizes the weights of all nodes to ensure that the sum of the weights is 1, and then calculates a weighted average of the feature extraction layer parameters according to the normalized weights to generate the next-generation global feature extraction model parameters.

[0079] After aggregation is complete, the central server distributes the newly generated global feature extraction model parameters to all participating nodes through a secure channel. This distribution process employs a differential version management mechanism, where each version of the global model has a unique identifier, facilitating version control and rollback operations for the nodes. Along with the parameters, the server also includes metadata information for this round of aggregation, such as the number of participating nodes, the weight distribution of each node, and the aggregation timestamp. This information helps nodes understand the global training status and make corresponding local decisions.

[0080] Step S330: Replace the feature extraction layer parameters of the corresponding localized model of each local node with the global model parameters of each local node, while retaining the local input adaptation layer and task-specific output layer parameters of the localized model.

[0081] After receiving the global feature extraction model parameters, each local node uses these parameters to replace the corresponding feature extraction layer parameters in its local model, thus updating the local model. This replacement operation ensures that all nodes maintain synchronization in feature extraction capabilities and share the feature representation capabilities enhanced by global knowledge. During the replacement process, each node retains its unique input adaptation layer and task-specific output layer parameters. These parameters contain node-specific data preprocessing logic and task solutions, representing the core of each subsystem's personalized capabilities.

[0082] This "globally shared + locally private" parameter management strategy achieves a balance between knowledge sharing and personalized needs. Taking the equipment monitoring subsystem as an example, the updated model possesses both the ability to recognize general equipment features learned from global data and retains anomaly detection logic specific to local equipment types. This design allows each node to benefit from the knowledge gains brought by federated learning while maintaining accurate adaptability to local specific tasks, forming an organic unity of personalization and universality.

[0083] Step S340: Repeat the above process until the global feature extraction model converges.

[0084] The complete process of training, uploading, aggregation, distribution, and updating described above will be repeated, forming a multi-round iterative federated learning loop. In each iteration, each node retrains locally based on the updated global feature extraction model, and then uploads the optimized feature extraction parameters again to participate in aggregation, and so on. The convergence state is determined by monitoring the changes in the performance metrics of the global model on the validation set. When the improvement of key metrics such as accuracy and F1 score in consecutive iterations is less than a preset threshold, the model is considered to have reached convergence.

[0085] Convergence determination employs a multi-indicator comprehensive evaluation mechanism. In addition to the traditional loss function value, it considers factors such as the model's performance stability across different categories and its robustness in marginal cases. This comprehensive evaluation method ensures that the final collaborative management model not only achieves optimal overall performance but also possesses reliable predictive capabilities in various sub-scenarios. Upon convergence, the final model obtained by each node integrates global knowledge and wisdom while retaining local advantages, forming a truly intelligent collaborative management and control system.

[0086] Further, step S310, which involves uploading the updated feature extraction layer parameters of the localized model to the central server, includes: Step S311: After homomorphic encryption or differential privacy processing is performed on the feature extraction layer parameters, the data is uploaded.

[0087] Before uploading the parameters from the feature extraction layer, local nodes employ homomorphic encryption or differential privacy techniques to enhance the privacy of the parameters. When homomorphic encryption is selected, a fully homomorphic encryption algorithm based on the RLWE problem is used to convert floating-point parameter values ​​into a polynomial representation in the encrypted state. This allows the central server to directly perform the arithmetic operations required for aggregation calculations on the encrypted parameters without decryption. Throughout the entire process, the parameters remain in ciphertext, effectively preventing privacy leaks during transmission and on the server side. During encryption, the numerical precision and structural integrity of the parameters are preserved, ensuring that the subsequent federated averaging algorithm can be executed correctly in the ciphertext space.

[0088] When employing a differential privacy protection scheme, precisely calibrated Laplacian or Gaussian noise is added to the feature extraction layer parameters based on the sensitivity analysis results. The noise scale parameter is determined jointly by the privacy budget ε and the global sensitivity of the parameters. This noise addition mechanism ensures that local data information of a single node cannot be inferred from the uploaded parameters, while statistically guaranteeing the accuracy of the aggregation results. After privacy processing, the parameters are uploaded to the central server via a secure HTTPS protocol, along with necessary metadata information for subsequent decryption or noise correction.

[0089] By introducing two complementary privacy enhancement techniques, homomorphic encryption and differential privacy, this method establishes a reliable security guarantee in the parameter uploading process. It not only achieves end-to-end encryption protection during data transmission but also provides a defense mechanism against potential privacy inference attacks, enabling the system to achieve effective collaborative learning while strictly protecting the data privacy of each mine subsystem.

[0090] Further, in step S320, the feature extraction layer parameters uploaded by all local nodes are received by the central server and aggregated using a federated averaging algorithm to obtain a global feature extraction model, including: Step S321: Based on the encrypted transmission of feature extraction layer parameters uploaded by each local node to the central server, the feature extraction layer parameters are verified for integrity and decrypted to obtain the plaintext feature extraction layer parameters of each local node.

[0091] The central server receives encrypted feature extraction layer parameters uploaded from each local node via a secure communication interface. These parameters are protected using an asymmetric encryption algorithm during transmission to ensure data security in intermediate transmission stages. The server first performs integrity verification on the received data packets, comparing digital signatures and hash values ​​to verify whether the parameters have been tampered with during transmission. Then, it uses a pre-set private key to decrypt the encrypted parameters, restoring the original feature extraction layer parameter matrix from each node. The decrypted parameters are temporarily stored in plaintext in the server's secure memory area for subsequent aggregation calculations, while a complete decryption log is recorded for audit trail purposes.

[0092] Step S322: Obtain the data volume factor, data quality factor, and model contribution factor of the local private data used for training by each local node in this training cycle, and calculate the plaintext feature extraction layer parameters of the corresponding local node and assign the corresponding aggregation weights.

[0093] In practice, after obtaining the plaintext feature extraction layer parameters of each local node, the central server initiates the composite weight calculation process. First, the system automatically counts the number of data samples actually used by each local node during the current training cycle, using this as the basis for calculating the data volume factor. Next, the system performs multi-dimensional quality assessment of the local private data: after converting the raw sensor data into feature vectors through the feature extraction layer, it calculates the average mutual information of these feature vectors to assess the statistical dependence between features, while simultaneously analyzing feature sparsity to determine the efficiency of feature representation; furthermore, it uses the isolated forest algorithm to perform anomaly detection on data batches, identifying abnormal samples caused by sensor malfunctions or environmental interference, and calculating the proportion of abnormal samples in the total data volume. Regarding model contribution evaluation, the system extracts the recent historical performance of the node from the maintained contribution decay memory queue, including the improvement in global model accuracy due to parameter updates in past training rounds, and calculates the current contribution factor using an exponentially weighted average algorithm, giving higher weight to recent high-quality contributions. Finally, the system weights and fuses the data volume factor, data quality factor, and model contribution factor, assigning higher weight coefficients to the data quality factor and model contribution factor to reflect the aggregation principle of "quality first, continuous optimization." In actual deployment, ventilation subsystem nodes typically achieve higher data quality factors due to the high stability of their sensor data; while equipment monitoring nodes, by continuously providing effective fault early warning model updates, often excel in model contribution factors. This composite weight calculation method ensures that during the collaborative training of multiple mine subsystems, the weight allocation of each node considers both immediate data quality and long-term contribution value.

[0094] Step S323: Normalize all aggregate weights to obtain the normalized weight value of each local node, and calculate the weighted average of the plaintext feature extraction layer parameters of all local nodes to obtain the global model parameters of the global feature extraction model.

[0095] All weights are normalized, converting the initial weights of each node into normalized weights that sum to 1. This is achieved by dividing the initial weight of each node by the sum of all weights. Subsequently, the normalized weights are used to calculate a weighted average of the feature extraction layer parameters for all nodes. Specifically, this involves summing the elements at each corresponding position of the parameter tensor according to their weights, generating the parameter matrix of the next-generation global feature extraction model. During the weighted average calculation, the integrity and dimensionality consistency of the parameter structure are simultaneously verified to ensure the correct execution of the aggregation operation.

[0096] Step S324: Construct a global feature extraction model based on global model parameters.

[0097] Based on the global model parameters calculated using a weighted average, a complete global feature extraction model is reconstructed, and the aggregated parameters are loaded into a neural network model with the same structure as the local nodes. The newly constructed global model is assigned a unique version identifier and records metadata information for this round of aggregation, including a list of participating nodes, the weight allocation for each node, and the aggregation timestamp. After construction is complete, basic integrity verification is performed on the global model to ensure the correct model structure and that the parameters have been effectively updated, preparing it for subsequent distribution and deployment.

[0098] Through a secure and reliable parameter receiving mechanism, intelligent weight allocation based on data volume, strict normalization processing, and accurate weighted average calculation, this method achieves effective integration of distributed knowledge while protecting the data privacy of all parties. The resulting new generation of global feature extraction model not only fully absorbs the local optimization experience of each node, but also maintains the uniformity and compatibility of the model structure, providing core technical support for building a continuously evolving smart mine collaborative management system.

[0099] Further, the data volume factor for obtaining the local private data used for training by each local node in this training cycle in step S322 includes: Step S3221a: Based on the amount of training data used by each local node in this training cycle, obtain the original data volume factor of the local node, where the amount of training data refers to the number of data samples actually used by the local node in this training.

[0100] In practice, each local node starts a data sampling counter at the beginning of each training cycle to count the number of valid data samples flowing through the model training process in real time. Taking the mine drainage subsystem as an example, its data acquisition system records monitoring parameters such as pump motor current, pipeline pressure, and water level at fixed sampling intervals. The multi-dimensional sensor readings at each sampling time point constitute a complete data sample. During local training, the integrity of each data sample is verified, and incomplete samples with sensor communication interruptions or data verification failures are removed to ensure counting accuracy. For the equipment monitoring subsystem, its vibration sensors collect equipment operating status data at a high frequency, which may generate tens of thousands of data samples in a single training cycle. However, the personnel positioning subsystem, due to the low update frequency of positioning tag data, typically only acquires a few thousand samples per cycle. This statistical method based on the actual number of data samples participating in the training can truly reflect the data processing scale of each subsystem in this round of training, providing a reliable basis for subsequent weight calculation.

[0101] Step S3221b: Normalize the original data volume factors of all local nodes to adjust the numerical range of each original data volume factor to a consistent scale, thus obtaining the data volume factor of each local node.

[0102] After completing the initial data volume statistics for all nodes, the weight calculation module of the central server initiates a normalization process. This module first iterates through all local nodes participating in this round of training, obtaining their respective initial data volume factors. Then, it maps these factors to a closed interval between zero and one using a Min-Max scaling algorithm. Specifically, it identifies the maximum and minimum initial data volume factors among all nodes, subtracts the minimum value from the initial data volume factor of each node, and divides the result by the difference between the maximum and minimum values ​​to obtain the normalized data volume factor. In actual deployment, when a subsystem's data volume is significantly lower due to equipment maintenance, this normalization process effectively prevents the node's data volume factor from being too small and excessively influencing the final weights. Simultaneously, for newly connected subsystem nodes, even with a small initial data volume, normalization ensures a reasonable weight participation, guaranteeing the stability of the federated learning system under dynamic node changes. The normalized data volume factors have a unified dimension and numerical range, laying the foundation for subsequent weighted calculations with data quality factors and model contribution factors.

[0103] Further, step S322, obtaining the data quality factor of the local private data used for training by each local node in this training cycle, includes: Step S3222a: Input the local private data of the local node into the feature extraction layer of the corresponding localization model to obtain the corresponding feature vector.

[0104] In the specific implementation process, each local node first converts the collected raw monitoring data into a standardized format suitable for model processing. Taking the ventilation subsystem as an example, its multi-channel wind speed sensors deployed in the roadway continuously generate time-series data. After preprocessing, these raw signals are formed into standard-length data segments, each containing wind speed and wind pressure monitoring values ​​at multiple consecutive time points. This preprocessed data is fed into the feature extraction layer of the localized model. This feature extraction layer typically consists of multiple convolutional and pooling layers, capable of extracting representative feature vectors from the raw time-series data. For the equipment monitoring subsystem, the raw waveform data collected by its vibration sensors is converted into frequency domain features through Fast Fourier Transform (FFT) before being input into the feature extraction layer to obtain a feature representation of the equipment status. This process maps the heterogeneous data from different subsystems to the same feature space, laying the foundation for subsequent quality assessment.

[0105] Step S3222b: Calculate the information metric of the feature vector. The information metric includes average mutual information or feature sparsity. Average mutual information is used to quantify the statistical dependence between feature vectors, and feature sparsity is used to quantify the distribution of non-zero elements in the feature vector.

[0106] After obtaining the feature vectors, two types of information metrics are calculated to evaluate feature quality. Average mutual information is calculated based on the statistical correlation between the dimensions of the feature vectors, quantifying the information content of the feature representation by estimating the difference between the joint distribution and marginal distribution among the feature dimensions. In practical applications of ventilation subsystems, feature vectors contain features across multiple dimensions such as wind speed stability and wind direction change patterns. A high average mutual information value indicates rich correlations between these features, effectively characterizing the operating status of the ventilation system. Feature sparsity assesses the compactness of features by calculating the proportion of significantly activated elements in the feature vectors. In equipment monitoring scenarios, feature vectors under normal operating conditions typically have high sparsity, while fault conditions lead to disordered feature activation patterns and reduced sparsity. These two metrics reflect the effectiveness of feature extraction from different perspectives, providing a quantitative basis for data quality assessment.

[0107] Step S3222c: Evaluate the ratio of outliers in the local private data. Outliers are identified using the Isolation Forest algorithm or an outlier detection method based on the confidence level of the local model prediction.

[0108] Anomaly detection algorithms are used to screen local private data for quality. The Isolation Forest algorithm identifies outliers that are significantly different from other samples by randomly partitioning the feature space into hyperplanes. This method is suitable for discovering anomalous data caused by sensor malfunctions or environmental interference. In the implementation of the drainage subsystem, this method analyzes the characteristic distribution of pump operating parameters and marks data that significantly deviate from the normal operating range as anomalous samples. Outlier detection based on local model prediction confidence uses a trained localized model to perform forward inference on the input data, classifying data samples whose model output confidence is lower than a set threshold as anomalous. In practical deployments, these two methods are often used in combination to improve the accuracy and recall of anomaly identification.

[0109] Step S3222d: Based on the information metric and the outlier ratio, the data quality factor is calculated by weighting. The information metric is positively correlated with the data quality factor, and the outlier ratio is negatively correlated with the data quality factor.

[0110] The final data quality factor is calculated based on the aforementioned evaluation results. Weighting coefficients are assigned to information metrics and outlier ratios, with average mutual information and feature sparsity given positive weights, and outlier ratios given negative weights. In practice, the ventilation subsystem, due to its more comprehensive sensor network, typically achieves higher scores on information metrics, while the equipment monitoring subsystem, whose vibration sensors are more susceptible to environmental interference, tends to have a higher outlier ratio. During the weighted calculation, the data characteristics of different subsystems are considered, and the weighting coefficients of each indicator are dynamically adjusted to ensure the fairness of the evaluation results. The final data quality factor accurately reflects the reliability and information value of the data from each subsystem.

[0111] Further, obtaining the model contribution factor of the local node in step S322 includes: Step S3223a: Obtain the historical contribution data of the local node in the past multiple rounds of aggregation from the contribution decay memory queue. The historical contribution data includes the instantaneous improvement of the performance index of the global model on the standard validation set after the feature extraction layer parameters of the local node are aggregated into the global model.

[0112] In the implementation process, the central server maintains a fixed-length contribution decay memory queue. This queue employs a first-in, first-out (FIFO) data management mechanism to continuously record the contribution performance of each local node in historical training cycles. After each round of federated aggregation, the performance of the updated global model is immediately evaluated on a standard validation set. This validation set contains representative, anonymized data samples collected from various mines, covering normal operating conditions and typical anomalies. Performance evaluation uses both accuracy and F1 score metrics. Accuracy reflects the overall prediction correctness, while the F1 score comprehensively considers precision and recall, making it particularly suitable for class imbalances common in mine data. The performance improvement resulting from the inclusion of each node's feature extraction layer parameters in the aggregation is precisely recorded. For example, parameter updates provided by the ventilation subsystem node may improve the accuracy of the global model in roadway wind speed prediction by a specific percentage, while the contribution of the equipment monitoring node is reflected in the increase in the F1 score for equipment fault identification. This historical contribution data is stored in the queue in chronological order, providing complete data support for subsequent contribution factor calculations.

[0113] Step S3223b: Based on historical contribution data, calculate the model contribution factor of the local node using the exponential weighted average method, where recent historical contributions are given higher weights.

[0114] Based on collected historical contribution data, an exponentially weighted average algorithm is used to calculate the model contribution factor for each node. This algorithm adjusts the weight of contribution values ​​from different periods by introducing a decay coefficient; recent contributions are given higher weights, while earlier historical contributions have their weights gradually decayed over time. In the implementation, a dynamically updated contribution factor value is maintained for each node. After each training round, the contribution factor from the current round is weighted and fused with the contribution factor from the previous round according to a preset ratio. For nodes requiring rapid response to changes in equipment status, such as the equipment monitoring subsystem, a larger decay coefficient is set to ensure that the contribution factor reflects the latest parameter update quality in a timely manner; while for nodes with relatively stable operating states, such as the ventilation subsystem, a smaller decay coefficient is used to ensure the stability of the contribution factor. This differentiated parameter configuration allows the contribution factor calculation to adapt to the operating characteristics of different subsystems and accurately reflect the continuous contribution capability of each node.

[0115] Step S3223c: Normalize the model contribution factors of all local nodes to adjust the numerical range of the model contribution factors to a consistent scale.

[0116] After calculating the model contribution factors for all nodes, a normalization process is initiated. This process first identifies the maximum and minimum contribution factor values ​​among all nodes, and then maps the contribution factors of each node to a uniform numerical range through a linear transformation. In practical applications, when a subsystem node provides high-quality parameter updates for multiple consecutive rounds, its contribution factor may be significantly higher than that of other nodes. Normalization effectively prevents a single node from excessively dominating the federated aggregation process. Simultaneously, for newly joined nodes or nodes with recent poor performance, normalization ensures they still receive basic participation weights, maintaining the inclusiveness of the federated learning system. The normalized contribution factors have consistent dimensions and distribution ranges, facilitating subsequent composite weight calculations with data quantity factors and data quality factors.

[0117] Further, step S322, which calculates the plaintext feature extraction layer parameters of the corresponding local node and assigns the corresponding aggregation weights, includes: Step S3224a: Based on the data volume factor, data quality factor and model contribution factor, the initial aggregation weight of each local node is calculated using the weighted geometric average method.

[0118] In the implementation process, the central server's weight calculation module receives data volume factors, data quality factors, and model contribution factors from each local node, and uses a weighted geometric mean algorithm to calculate the initial aggregate weights. This algorithm ensures that any factor approaching zero will significantly suppress the final result, effectively avoiding the negative impact of low-quality data or negative nodes on the global model. In actual deployment, different weight coefficients are configured for each of the three factors, with data quality factors and model contribution factors typically assigned higher weight proportions to reflect the principles of quality priority and continuous contribution. Taking the ventilation subsystem as an example, when it has both a high data quality factor and a stable model contribution factor, it can still obtain a high initial aggregate weight even if the data volume factor is relatively low; conversely, if the data quality factor of a node drops significantly due to sensor failure, even if its data volume is large and its historical contribution is good, its initial weight will be significantly suppressed. This calculation method can automatically identify and prioritize high-quality, high-value parameter updates, laying the foundation for subsequent weight adjustments.

[0119] Step S3224b: Based on the safety risk level of the mining subsystem corresponding to the local node in the current operating cycle, the initial aggregation weight is dynamically adjusted. The safety risk level is positively correlated with the initial aggregation weight.

[0120] After obtaining the initial aggregation weights, the weights are dynamically adjusted based on real-time risk assessment data provided by the mine safety production monitoring platform. The safety risk level is calculated comprehensively from multiple indicators, including methane concentration monitoring values ​​for each area, personnel distribution density, and equipment operating status index. This data is transmitted to the central server in real time via the mine's industrial ring network. When the safety risk level of a certain area increases, the weights of subsystem nodes related to that area receive a corresponding increase coefficient. For example, the weight increase coefficient for the ventilation subsystem increases when methane concentration exceeds limits, and the weight increase coefficient for the personnel positioning subsystem increases when personnel gathering exceeds limits. In actual operation, a complete mapping relationship between risk levels and weight adjustment coefficients has been established: low-risk states correspond to a baseline adjustment coefficient, medium-risk states have moderately increased coefficients, and high-risk states use the maximum increase coefficient. This dynamic adjustment mechanism ensures that the federated learning process can respond quickly to changes in the mine safety situation, prioritizing the model update needs of key areas.

[0121] Step S3224c: Normalize the aggregate weights of all local nodes after dynamic adjustment to obtain the aggregate weight values ​​used for weighted geometric mean calculation.

[0122] After dynamic weight adjustment, the aggregated weights of all nodes are normalized. This process first calculates the sum of all weights, then divides the weight of each node by this sum to ensure that the final sum of all weights is exactly one. In the implementation, double-precision floating-point arithmetic is used to guarantee calculation accuracy and avoid cumulative bias caused by rounding errors. For example, when multiple subsystems are simultaneously in a high-risk state and their weights are all increased, normalization can reasonably balance the weight ratios among nodes, preventing a single node from excessively dominating the model aggregation process. Simultaneously, intermediate calculation results during the normalization process are recorded, including statistical information such as the total weight and the maximum weight value, for subsequent performance analysis and system optimization. The normalized aggregated weights have uniform dimensions and standardized numerical ranges, and can be directly used for subsequent weighted average calculations.

[0123] In one specific embodiment of the present invention, the mining subsystem is an independently operating and data-heterogeneous subsystem in a smart mine, including at least one of personnel positioning, ventilation, drainage and equipment monitoring.

[0124] The personnel positioning subsystem collects and processes real-time information such as three-dimensional coordinates, movement trajectories, and dwell time in different areas using UWB positioning base stations deployed in the roadways and positioning tags carried by personnel. This subsystem employs a deep learning-based trajectory prediction model, inputting continuous positioning coordinate sequences into the feature extraction layer of a localized model. A temporal convolutional network extracts the spatiotemporal features of personnel movement, and a task-specific output layer accurately classifies the personnel's work area or predicts their future movement trajectory, providing data support for mine personnel safety management and emergency evacuation. The ventilation subsystem continuously monitors the mine's ventilation network operation using wind speed sensors, wind pressure sensors, and air door status sensors distributed in key areas such as the main roadway and working face. The localized model built by this subsystem uses multi-dimensional sensor time-series data as input, learns the dynamic characteristics and normal operating modes of the ventilation system through the feature extraction layer, predicts airflow demand at key nodes through the regression output layer, or identifies abnormal ventilation states through the classification output layer, providing a basis for intelligent control and energy consumption optimization of the ventilation system. The drainage subsystem integrates pump motor current sensors, water level monitoring sensors, and pipeline pressure sensors to collect real-time data on the operating status of drainage equipment and changes in mine water level. The localized model of this subsystem employs a multi-source data fusion architecture, aligning asynchronously sampled equipment status data with continuous water level data and extracting features. It predicts mine water level change trends through regression tasks or identifies early failures of pump equipment through anomaly detection tasks, ensuring reliable operation and preventative maintenance of the drainage system under complex conditions. The equipment monitoring subsystem collects multimodal monitoring data during equipment operation through vibration acceleration sensors, temperature sensors, and sound sensors installed on key equipment such as mining and transportation equipment. The localized model of this subsystem uses a specially designed feature extraction network to process the frequency domain features of vibration signals, the trend features of temperature signals, and the time-frequency features of sound signals. The output layer of the autoencoder structure calculates the reconstruction error of the equipment status, enabling early detection and warning of abnormal equipment conditions, providing technical support for predictive maintenance of mining equipment.

[0125] Accordingly, a second aspect of the present invention provides an electronic device, including: at least one processor; and a memory connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to execute the above-described intelligent mine multi-subsystem collaborative management method based on federated transfer learning.

[0126] Accordingly, a third aspect of the present invention provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the above-described intelligent mine multi-subsystem collaborative management method based on federated transfer learning.

[0127] This invention aims to protect a collaborative management method for multi-subsystem smart mines based on federated transfer learning, which has the following advantages: 1. By using a federated learning framework, each mine subsystem can effectively integrate knowledge from isolated subsystems such as personnel positioning, ventilation, and drainage without uploading local private raw data. This is achieved by exchanging and aggregating encrypted intermediate layer feature vectors, thus training an enhanced model with a global perspective. This breaks the inherent "data silo" dilemma. When optimizing the ventilation system, personnel positioning information can be comprehensively considered. When diagnosing equipment faults, the operating status of the drainage system can be correlated. Ultimately, this significantly improves the accuracy and globality of intelligent decision-making tasks such as safety risk warning and production process optimization. 2. The system adopts a mechanism of "centralized pre-training combined with hierarchical federated collaboration". During the federated iteration process, each local node only needs to upload the feature extraction layer parameters of the model to the central server, while the massive amount of raw data and the local private input / output layer parameters are kept locally. This extreme compression of the amount of uploaded data reduces the total communication load of the system from the terabit level of traditional centralized learning or full-parameter federated learning to the gigabit level, which greatly alleviates the bandwidth pressure on the mining industrial ring network and enables the system to access and expand more subsystems at a lower cost. 3. Through the model architecture design of "shared feature layer + private task layer", supplemented by pre-training initialization and phased training strategies, the final collaborative model has two advantages: on the one hand, the shared feature extraction layer obtained by federated aggregation optimization integrates the common knowledge of all subsystems and has a powerful and universal feature representation capability; on the other hand, the local input adaptation layer and task-specific output layer retained by each subsystem enable it to accurately adapt to its own unique data structure and management task, thereby ensuring the high performance of the global model while perfectly meeting the personalized needs of different subsystems and achieving precise management of "harmony in diversity".

[0128] 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.

[0129] 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.

[0130] 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.

[0131] 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.

[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A wisdom mine multi-subsystem collaborative management method based on federal transfer learning, characterized in that, The collaborative management system comprises a central server and local nodes corresponding to a plurality of mine subsystems, and the method comprises the following steps: Based on the central server, the deep neural network is pre-trained based on the industrial equipment data set and the historical desensitization data to obtain an initial global feature extraction model, and the global model parameters of the initial global feature extraction model are distributed to a plurality of local nodes; Based on each local node receiving the global model parameters, a localized model is constructed, and the local private data of the local node is used to locally adaptively train the localized model; Based on the central server and all the local nodes, collaborative training is performed through a federated learning framework, each local node uploads the feature extraction layer parameters obtained after adaptive training of the localized model to the central server, the central server aggregates the feature extraction layer parameters uploaded by all the local nodes, updates the current global feature extraction model, generates a new global feature extraction model and distributes it to each local node, and the process is repeated until the model converges, thereby obtaining a collaborative management model of the collaborative management system based on the evolved global feature extraction model.

2. The smart mine multi-subsystem coordination management method based on federal transfer learning according to claim 1, characterized in that, The collaborative training based on the central server and all the local nodes through the federated learning framework, each local node uploading the feature extraction layer parameters obtained after adaptive training of the localized model to the central server, the central server aggregating the feature extraction layer parameters uploaded by all the local nodes, updating the current global feature extraction model, generating a new global feature extraction model and distributing it to each local node, and repeating the process until the model converges, comprises: Based on each local node, the corresponding localized model is trained and updated based on the local private data, and the feature extraction layer parameters of the updated localized model are uploaded to the central server; Based on the central server, all the feature extraction layer parameters uploaded by the local nodes are received, and a federated averaging algorithm is used for aggregation to obtain a global feature extraction model, and the global model parameters of the global feature extraction model are distributed to all the local nodes; Based on the global model parameters of each local node, the feature extraction layer parameters of the corresponding localized model of the local node are replaced, while the local input adaptation layer and task-specific output layer parameters of the localized model are retained; The above process is repeated until the global feature extraction model converges. 3.The smart mine multi-subsystem coordination management method based on federal transfer learning according to claim 2, characterized in that, The global feature extraction model obtained by the central server aggregating the feature extraction layer parameters uploaded by all the local nodes through the federated averaging algorithm, comprises: Based on the central server, the encrypted transmission of the feature extraction layer parameters uploaded by each local node is received, and integrity verification and decryption processing are performed on the feature extraction layer parameters to obtain the plaintext feature extraction layer parameters of each local node; obtaining a data quantity factor, a data quality factor and a model contribution factor of local private data of each local node used for training in the current training period, calculating an aggregation weight of the plaintext feature extraction layer parameter of the corresponding local node according to the data quantity factor, the data quality factor and the model contribution factor; normalizing all the aggregation weights to obtain a normalized weight value of each local node, and performing weighted average calculation on the plaintext feature extraction layer parameters of all the local nodes to obtain global model parameters of the global feature extraction model; constructing the global feature extraction model based on the global model parameters.

4. The federated transfer learning based intelligent mine multi-subsystem collaborative management method according to claim 3, characterized in that, The data quantity factor of the local private data of each local node used for training in the current training period is obtained, including: obtaining an original data quantity factor of each local node based on the training data quantity used by the local node in the current training period, wherein the training data quantity refers to the number of data samples actually participating in model training of the local node in the current training; normalizing the original data quantity factors of all local nodes to adjust the numerical value range of each original data quantity factor to a consistent scale to obtain the data quantity factor of each local node.

5. The federated transfer learning based intelligent mine multi-subsystem collaborative management method according to claim 3, characterized in that, The data quality factor of the local private data of each local node used for training in the current training period is obtained, including: inputting the local private data of the local node into the feature extraction layer of the corresponding local model to obtain a corresponding feature vector; calculating an information metric of the feature vector, the information metric including average mutual information or feature sparsity, wherein the average mutual information is used to quantify the statistical dependence between feature vectors, and the feature sparsity is used to quantify the distribution of non-zero elements in the feature vector; evaluating the ratio of abnormal samples in the local private data, the abnormal samples being identified by an isolation forest algorithm or an outlier detection method based on local model prediction confidence; obtaining the data quality factor by weighted calculation based on the information metric and the abnormal sample ratio, wherein the information metric is positively correlated with the data quality factor, and the abnormal sample ratio is negatively correlated with the data quality factor.

6. The federated transfer learning based intelligent mine multi-subsystem coordination management method according to claim 3, characterized in that, The model contribution factor of the local node is obtained, including: obtaining historical contribution data of the local node in past multiple rounds of aggregation from a contribution decay memory queue, the historical contribution data including an instantaneous performance improvement amplitude of the global model on a standard validation set after the feature extraction layer parameters of the local node are aggregated to the global model; calculating the model contribution factor of the local node by using an exponential weighted average method based on the historical contribution data, wherein recent historical contribution is given a higher weight; normalizing the model contribution factors of all the local nodes to adjust the numerical value range of the model contribution factors to a consistent scale.

7. The federated transfer learning based intelligent mine multi-subsystem collaborative management method according to claim 3, characterized in that, The calculation of the aggregation weight of the plaintext feature extraction layer parameter of the corresponding local node includes: calculating an initial aggregation weight of each local node by using a weighted geometric average method based on the data quantity factor, the data quality factor and the model contribution factor; Based on the safety risk level of the mining subsystem corresponding to the local node in the current operating cycle, the initial aggregation weight is dynamically adjusted, and the safety risk level is positively correlated with the initial aggregation weight. The aggregate weights of all the dynamically adjusted local nodes are normalized to obtain the aggregate weight values ​​used for calculating the weighted geometric mean.

8. The federated transfer learning based intelligent mine multi-subsystem collaborative management method according to any one of claims 1-7, characterized in that, The step of receiving the global model parameters from each local node, constructing a localized model, and using the local private data of each local node to perform local adaptation training on the localized model includes: Based on the global model parameters of the initial global feature extraction model, a local base model with the same network structure is constructed on the local node. Based on the local base model, an input adaptation layer for adapting to local data dimensions and a task-specific output layer configured according to the local management task type are added to obtain the localized model. While keeping the model parameters of the local base model frozen, the local private data of the local node is used to train the input adaptation layer and the task-specific output layer in the localized model until the loss function of the newly added network layer reaches a stable state, thus completing the local adaptation training. 9.The smart mine multi-subsystem coordination management method based on federated transfer learning according to claim 8, characterized in that, While keeping the model parameters of the local base model frozen, the local private data of the local node is used to train the input adaptation layer and the task-specific output layer in the localized model until the loss function of the newly added network layer reaches a stable state, including: A forward propagation computation process is constructed on the local node, and the local private data is input into the localized model. The model then passes through the local base model in a parameter-frozen state, the input adaptation layer in a trainable state, and the task-specific output layer in sequence to obtain the model prediction output. Calculate the difference between the model's predicted output and the true label, and calculate the loss value for the current training state based on this difference using a loss function; The backpropagation algorithm is executed, and the gradient update of the backpropagation algorithm is limited to the model parameters of the input adaptation layer and the task-specific output layer; Based on the gradient calculated by the backpropagation algorithm, the model parameters of the input adaptation layer and the task-specific output layer are updated using the stochastic gradient descent algorithm. Repeat the above training process until the change in the loss function value is less than the preset convergence threshold in multiple consecutive training cycles, thus completing the training of the new network layer.

10. The federated transfer learning based intelligent mine multi-subsystem coordination management method according to claim 8, characterized in that, The task-specific output layer configured according to the local management task type includes: Linear output layer for regression tasks; A Softmax output layer used for classification tasks; or, An autoencoder output structure for anomaly detection tasks.

Images