Adaptive variable weight coal rock dynamic disaster comprehensive evaluation method

Abstract

The application discloses a kind of self-adapting variable weight coal rock dynamic disaster comprehensive evaluation method, it is related to coal mine safety monitoring technical field, for the evaluation problem caused by the difference of multiple mining area geology and equipment in coal mine enterprise intelligent monitoring network, traditional fixed weight method is difficult to adapt to regional difference, often cause false alarm and dispatch instability;Through local data processing and cloud cooperation realizes dynamic, individualized evaluation: mining area node carries out calibration screening to sensor data, trains self-adapting weight network to generate initial weight;Cloud aggregation encrypts update data, constructs weight manifold and calculates geological embedding vector, generates individualized transformation matrix;Mining area node generates dispatch priority in combination with transformation matrix and risk signal;Cloud continues to feed back and optimize weight manifold.Compared with traditional method, the application significantly improves evaluation accuracy, stability and real-time response capability, enhances cross-domain collaborative scheduling feasibility, effectively guarantees mining area safety and production efficiency.

Description

Technical Field

[0001] This invention relates to the field of coal mine safety monitoring technology, specifically to an adaptive variable weight coal and rock dynamic disaster comprehensive assessment method. Background Technology

[0002] Coal mining companies are deploying intelligent monitoring networks spanning multiple work areas. Underground sensors continuously collect information on earthquakes, surrounding rock stress, gas content, and temperature and humidity conditions, which is then aggregated to the cloud via dedicated communication links. Due to significant differences in geological age, fracture development, burial depth, and equipment configuration across different mining areas, the same monitoring indicators can represent different disaster mechanisms and impact intensities at different locations. The dispatch center aims to use a unified comprehensive assessment platform for real-time risk ranking and coordinated control. However, traditional methods often rely on prior experience to assign a fixed set of weights, which are then applied to all mining areas. As the scale of data flow expands and regional differences become more apparent, this "one-size-fits-all" approach is gradually revealing problems such as insufficient sensitivity to local characteristics and unstable assessment results, thus weakening the feasibility of cross-mining area collaborative dispatch.

[0003] The fundamental reason for the cross-domain failure of the fixed-weight model is: A subtle but persistent shift in weighting patterns has emerged between different mining areas: in one roadway, gas levels might be the primary trigger, while in another, surrounding rock stress might dominate. When the assessment system directly transfers the original weights to the new scenario, it's like misplacing the risk "scale" on a mismatched scale. This can either amplify warning signals, leading to frequent false alarms, or mask real hazards, causing missed intervention opportunities. Over time, this mismatch can create oscillations between the cloud model and on-site feedback, resulting in frequent revisions of dispatch instructions and confusion among on-site personnel. To truly implement the "Adaptive Variable Weight Coal and Rock Dynamic Disaster Comprehensive Assessment Method," each mining area must first learn its own most accurate weight distribution locally. Then, through federated collaboration, encrypted gradients or statistics should be shared to the cloud. Spatial alignment and continuous fine-tuning should be achieved using weight mapping and geological attribute embedding. This ensures that the comprehensive risk index respects regional differences while maintaining platform comparability, thereby tightly integrating intelligent assessment, collaborative decision-making, and dynamic early warning. Summary of the Invention

[0004] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an adaptive variable-weight comprehensive assessment method for coal and rock dynamic hazards. This method achieves dynamic and personalized assessment through local data processing and cloud collaboration: mining nodes calibrate and filter sensor data, training an adaptive weighted network to generate initial weights; the cloud aggregates and encrypts updated data, constructs a weighted manifold, calculates geological embedding vectors, and generates a personalized transformation matrix; mining nodes combine the transformation matrix and risk signals to generate scheduling priorities; and the cloud continuously provides feedback to optimize the weighted manifold. Compared to traditional methods, this invention significantly improves assessment accuracy, stability, and real-time response capabilities, enhances the feasibility of cross-domain collaborative scheduling, and solves the technical problems described in the background section.

[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: An adaptive variable weight coal and rock dynamic disaster comprehensive assessment method includes: each mining area performs synchronous calibration and credibility screening of underground multi-source sensor streams locally, generates standardized feature tensors and calculates the corresponding encrypted gradient statistics; Mining nodes train a lightweight adaptive weighted network using standardized feature tensors as input, obtain an initial weight vector, and package it together with encrypted gradient statistics to the cloud. After receiving the encrypted update from multiple mining areas, the potential weight manifold is obtained and the geological structure embedding vector is calculated for each mining area; Based on the geological structure embedding vector, the weight vector of each mining area is mapped to the potential weight manifold, generating a personalized transformation matrix and back to the mining area node; After receiving the transformation matrix, the mining area node immediately calculates the comprehensive disaster value and generates a real-time scheduling priority coefficient by jointly quantifying the risk growth signal and the response lag signal. Based on this, control commands are issued and performance feedback indicators are synchronized to the cloud. Based on accumulated performance feedback metrics and geological drift indicators, the cloud periodically updates the potential weight manifold, regenerates the transformation matrix, and synchronizes it to all mining area nodes.

[0006] Furthermore, the multi-source sensor streams from underground are processed locally in the mining area. The multi-source sensor data includes: collected seismic data, surrounding rock stress, gas content, and temperature and humidity data. The raw calibration dataset is generated through time alignment and measurement calibration. Based on historical data and sensor performance characteristics, sliding window analysis and isolated forest algorithms are used to identify anomalous data points and remove unreliable data to generate a reliable dataset.

[0007] Furthermore, time-domain features, frequency-domain features, and spatial features are extracted from the trusted dataset, and standardized to eliminate dimensional differences, and organized into standardized feature tensors. Homomorphic encryption technology is used to encrypt the standardized feature tensors to generate encrypted feature tensors, and encrypted gradient statistics are calculated based on the encrypted feature tensors.

[0008] Furthermore, during the training of the lightweight adaptive weighted network, the mining nodes record the gradient changes of the network parameters, calculate the mean gradient and the magnitude of gradient changes to generate gradient statistics, and encrypt the gradient statistics to form encrypted gradient statistics.

[0009] Furthermore, encrypted update data from multiple mining areas is received, including encrypted gradient statistics and initial weight vectors. The encrypted update data is decrypted using a preset decryption key to generate decrypted gradient statistics and initial weight vectors. Manifold learning is applied to the initial weight vectors to generate potential weight manifolds through dimensionality reduction and clustering, which characterize the similarity and difference of weight distributions between mining areas.

[0010] Furthermore, the geological structure data, including geological age, fault development degree and burial depth, are processed to generate geological feature vectors. An autoencoder is then used to convert the geological feature vectors into low-dimensional geological structure embedding vectors. The cosine similarity between the geological structure embedding vector and the corresponding mining area coordinates in the potential weight manifold is calculated to quantify the influence of geological features on the weight distribution.

[0011] Furthermore, the initial weight vectors of each mining area are mapped to the potential weight manifold using the geological structure embedding vector, and the weight space is aligned by calculating the personalized transformation matrix.

[0012] Furthermore, the personalized transformation matrix is ​​generated by minimizing the squared Euclidean distance between the mapped weight vector and the mining area coordinates in the latent weight manifold, and is fine-tuned and optimized based on the cosine similarity between the geological structure embedding vector and the mining area coordinates. The optimized personalized transformation matrix is ​​sent back to the mining area node, which then uses this personalized transformation matrix to adjust its local weight vector for subsequent calculation of comprehensive disaster values.

[0013] Furthermore, after receiving the personalized transformation matrix, the mining node uses the personalized transformation matrix and the standardized feature tensor to calculate the comprehensive disaster value. It generates a real-time scheduling priority coefficient through the joint quantization of the risk growth signal and the response lag signal, and sorts and issues control commands according to the real-time scheduling priority coefficient. At the same time, it uploads the performance feedback indicators.

[0014] Furthermore, performance feedback indicators and geological drift signs of each mining area node are collected, a performance feedback matrix is ​​constructed based on the performance feedback indicators, and a drift sign vector is constructed based on the geological drift signs. The comprehensive performance score and influence factor are calculated using the performance feedback matrix and drift indicator vector. The potential weight manifold is updated using the comprehensive performance score and influence factor. Based on the updated potential weight manifold, a personalized transformation matrix is ​​generated for each mining node. The personalized transformation matrix is ​​transmitted to the corresponding mining node through a dedicated communication link.

[0015] Preferred, (III) Beneficial Effects This invention provides an adaptive variable weight comprehensive assessment method for coal and rock dynamic hazards, which has the following beneficial effects: This solution achieves dynamic weight adjustment through a local adaptive weight network and a continuous cloud feedback mechanism. Traditional methods use fixed weights, which struggle to cope with long-term evolutions in geological conditions, environmental changes, or equipment status, leading to assessment results that gradually deviate from reality. This method, however, utilizes a lightweight weight network trained in real-time on mining nodes, combined with a weight manifold updated in the cloud based on geological drift signals. This allows the assessment model to quickly adapt to changes, ensuring the timeliness and applicability of the results. This feature significantly improves the system's performance in complex and ever-changing mining environments.

[0016] To address the differences in geological age, fracture development, burial depth, and equipment configuration across various mining areas, this scheme introduces geological structure embedding vectors and personalized transformation matrices. Each mining area node generates an initial weight vector based on local multi-source sensor data. The cloud platform then adjusts the weights to personalized parameters tailored to the specific characteristics of each mining area through cross-domain data aggregation and mapping. Compared to the traditional "one-size-fits-all" uniform weighting model, this method significantly improves the sensitivity of assessments to local features, enhances the accuracy of disaster prediction, and reduces misjudgments caused by regional differences.

[0017] This solution employs federated data sharing and cloud-based weight mapping technology to achieve collaborative assessment across multiple mining areas while protecting sensitive information in each area. Traditional methods, when applied across mining areas, struggle to balance uniformity and individual needs due to a lack of regional differentiation processing. This method, however, maintains the comparability of assessment results while respecting the uniqueness of each mining area through encrypted gradient statistical uploading and cloud aggregation. This significantly improves the feasibility of cross-domain collaborative scheduling and provides strong support for enterprise-level risk management.

[0018] In the data processing stage, this solution ensures the quality of input data through local calibration and reliability screening. In the weight optimization stage, adaptive networks and feedback mechanisms effectively reduce false alarms and missed alarms. Traditional fixed-weight methods often amplify noise signals or mask potential hazards due to weight mismatch, while this method, through dynamic adjustment and continuous optimization, makes the evaluation results more stable and reliable. This improvement directly enhances the accuracy of disaster early warning and provides a more solid guarantee for mine safety.

[0019] This solution incorporates a mechanism for rapidly generating scheduling priorities. After receiving the transformation matrix from the cloud, the mining nodes combine it with risk acceleration and response lag signals to calculate the overall disaster magnitude and priority coefficient in real time. Compared to the lag inherent in traditional methods, this approach enables a rapid response in the early stages of potential threats, shortening warning and response times, providing a valuable window for disaster relief and evacuation, and significantly improving the efficiency of emergency management. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the adaptive variable weight coal and rock dynamic disaster comprehensive assessment method of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1 This invention provides an adaptive variable weighted comprehensive assessment method for coal and rock dynamic hazards, comprising: Step 1: Process the multi-source sensor data stream locally in the mining area. This includes: collecting multi-source sensor data such as seismic, surrounding rock stress, gas content, and temperature and humidity data; generating a raw calibration dataset through time alignment and measurement calibration; establishing normal operating ranges based on historical data and sensor performance characteristics; identifying abnormal data points using sliding window analysis and isolated forest algorithms; removing unreliable data to generate a reliable dataset; extracting time-domain, frequency-domain, and spatial features from the reliable dataset; standardizing the dataset to eliminate dimensional differences; and organizing the data into a standardized feature tensor; encrypting the standardized feature tensor using homomorphic encryption technology to generate an encrypted feature tensor; and calculating encrypted gradient statistics based on the encrypted feature tensor to protect data privacy and support subsequent cloud analysis.

[0023] Step 101: Sensor Data Acquisition and Synchronous Calibration In the underground mining environment, multiple sensors are deployed to collect environmental data in real time, forming a multi-source sensor data stream. These sensors include seismic sensors, stress sensors, gas sensors, and temperature and humidity sensors, which are used to monitor key indicators such as seismic activity, surrounding rock stress, gas content, and temperature and humidity. The collected data is recorded in time series format, including information such as seismic amplitude, surrounding rock stress value, gas concentration, and temperature and humidity.

[0024] To ensure data consistency in time and measurement accuracy, the timestamps of all sensor data are first aligned. This is done by adjusting the timestamps of each sensor's data using the sensor's built-in clock or an external unified clock source (such as the GPS time signal) to align them on the time axis. Secondly, calibration is performed for the measurement characteristics of different sensors. For example, zero-point calibration is performed on the stress sensor to eliminate initial offset, and temperature drift correction is performed on the temperature and humidity sensor to compensate for the effects of environmental changes. After time alignment and measurement calibration, the original calibration dataset is generated.

[0025] By aligning the data with time and calibrating the measurement, inconsistencies caused by sensor time deviations or measurement errors are eliminated, ensuring the consistency of multi-source sensor data in terms of time series and numerical accuracy. This processing method improves the reliability and accuracy of the data, providing a reliable input for subsequent data analysis and processing.

[0026] Step 102, Credibility Screening Based on the original calibration dataset, unreliable data affected by noise, sensor malfunctions, or external interference is further removed to improve data quality. The specific method includes the following steps: First, based on historical data and sensor performance characteristics, establish normal operating ranges for each type of sensor. For example, the concentration values ​​of gas sensors are usually within a predefined reasonable range, and data that exceeds this range or shows abrupt changes are considered potential anomalies.

[0027] Secondly, a sliding window analysis technique is used to perform local statistical calculations on the time series data, analyze the degree of deviation of each data point from the surrounding data, such as comparing its value with the local mean or median, and combine it with the isolated forest algorithm to identify outlier data points.

[0028] Finally, each data point is assigned a confidence score between 0 and 1, where 1 represents complete confidence and data points with a score less than 0.5 are considered unreliable and are removed. After filtering, a reliable dataset is formed.

[0029] By establishing a normal working range, using sliding window analysis, and assigning confidence values, abnormal data can be effectively identified and removed, thereby ensuring the quality and reliability of the trusted dataset. This screening process reduces the interference of noisy or faulty data on subsequent analysis and enhances the stability and accuracy of the system.

[0030] Step 103: Generation of Standardized Feature Tensors Based on a trusted dataset, multi-dimensional features are extracted and standardized to generate feature tensors suitable for subsequent analysis. The specific process is as follows: First, time-domain features of the time-series data are extracted, such as calculating the peak value, mean, and rate of change, where the rate of change reflects the dynamic trend of gas concentration or stress value. Second, a Fast Fourier Transform (FFT) is performed on the seismic data to extract frequency-domain features, including the dominant frequency component and energy spectrum characteristics, to analyze the frequency distribution of the seismic signal. Third, spatial correlation coefficients are calculated based on the spatial location of the sensors within the mining area, reflecting the distribution characteristics of surrounding rock stress in different regions. Next, all extracted features are normalized by subtracting the mean of each feature value and then dividing by its standard deviation to eliminate dimensional differences between features, resulting in standardized feature values. Finally, the standardized feature values ​​are organized into a three-dimensional tensor according to sensor type, time series, and spatial dimension. The structure of this tensor is composed of the number of sensors, time step, and feature dimension, generating a standardized feature tensor.

[0031] By extracting time-domain, frequency-domain, and spatial features, multi-dimensional information about the downhole environment is comprehensively captured. Standardization eliminates the dimensional differences between feature values, enabling different types of data to be compared and analyzed on a uniform scale. The generated three-dimensional tensor structure provides a systematic and efficient data input format for subsequent network training, improving the adaptability and accuracy of data processing.

[0032] Step 104: Encryption Gradient Statistical Calculation To protect the privacy of local data in the mining area and support subsequent computational needs, the standardized feature tensor is encrypted and statistical information is calculated. The specific steps are as follows: First, homomorphic encryption is used to encrypt the standardized feature tensor, generating an encrypted feature tensor. This encryption method allows mathematical operations to be performed in the encrypted state. Second, during the initial training of the local adaptive weighted network, initial gradients are calculated based on the standardized feature tensor. Specifically, this involves analyzing the direction and magnitude of the feature tensor's changes, without needing to complete the entire network training. Then, gradient calculations are performed on the encrypted feature tensor within the encrypted domain to obtain the encrypted gradient. Finally, statistical features of the encrypted gradient are extracted, such as calculating the mean and quantiles of the encrypted gradient magnitude, generating encrypted gradient statistics. These statistics will be uploaded to the cloud for subsequent processing.

[0033] Homomorphic encryption technology ensures the privacy and security of data during local processing and transmission, while supporting computational operations within the encrypted domain, thus avoiding the risk of data leakage. Extracting encrypted gradient statistics provides essential data support for cross-domain collaboration in the cloud, enabling collaborative analysis of data from different mining areas while protecting privacy, thereby improving the system's security and usability.

[0034] This invention completes the processing flow from multi-source sensor data to high-quality encrypted statistical information through four sub-steps: sensor data acquisition and synchronous calibration, reliability screening, standardized feature tensor generation, and encrypted gradient statistical calculation. Each sub-step is interconnected, first ensuring the time and measurement accuracy of the data, then improving data quality, followed by feature extraction and standardization, and finally generating statistical information under privacy protection. This technical logic not only improves the accuracy and reliability of the data but also enhances the system's security through privacy protection measures, providing solid data support for subsequent local network training and cloud analysis, demonstrating the technical advantages of this invention in data processing and security protection.

[0035] Step 2: During the training of the lightweight adaptive weighted network, the mining nodes record the gradient changes of the network parameters, calculate the mean gradient and the magnitude of gradient changes to generate gradient statistics, and use homomorphic encryption technology to encrypt the gradient statistics to form encrypted gradient statistics. Step 201: Construction and Training of a Lightweight Adaptive Weighted Network A lightweight neural network was constructed on local equipment in the mining area to learn the weight distribution of monitoring indicators. This network consists of three parts: an input layer, hidden layers, and an output layer. The input layer receives a normalized feature tensor, which is a composite of multi-source sensor data after time synchronization calibration and confidence filtering. The hidden layer employs a multilayer perceptron structure, processing the input data through multiple neurons to capture the nonlinear correlation characteristics between monitoring indicators. The output layer generates an initial weight vector with dimensions consistent with the number of monitoring indicators. To accommodate the limited computing power of local equipment in the mining area, the network is designed to be lightweight, reducing computational complexity and resource consumption.

[0036] The goal of the training process is to make the comprehensive risk index as close as possible to the actual occurrence of disasters.

[0037] The comprehensive risk index is defined as the weighted sum of monitoring indicators and their corresponding weights. Historical disaster data is used as a reference standard during training. By repeatedly adjusting network parameters, the initial weight vector is made to reflect the unique risk distribution characteristics of the mining area. Parameter adjustment is achieved using the Adam optimizer, with a fixed learning rate of 0.001 and a fixed number of training epochs of 100 to ensure a balance between convergence speed and computational efficiency. To prevent the model from becoming too closely aligned with the training data and losing its generalization ability, an L2 regularization term is added to the loss function. This improves the model's stability by limiting excessively large weight values. After training, the network outputs an initial weight vector, serving as the basis for local risk assessment and providing input for subsequent cloud-based collaboration.

[0038] The lightweight neural network can be built to run efficiently on local equipment in mining areas with limited computing resources, quickly generate initial weight vectors to meet real-time requirements, and use historical disaster data to guide training to ensure that the initial weight vectors can accurately reflect the risk characteristics of the local mining area and improve the pertinence of risk assessment. The Adam optimizer and L2 regularization term are introduced to optimize the training process and ensure the stability and reliability of the model in complex environments.

[0039] Step 202: Generation of encrypted gradient statistics During the training of a lightweight adaptive weighted network, the gradient changes of the network parameters are recorded, and statistical information is extracted from them. The gradient statistics include two parts: the mean gradient and the magnitude of the gradient change. The gradient mean is obtained by averaging the gradients of all network parameters, reflecting the overall trend of gradient changes. The gradient magnitude is obtained by first calculating the deviation of each gradient from the mean, then averaging the sum of squares of all deviations, and finally taking the square root of the result, reflecting the dispersion of the gradient distribution. These two statistics together constitute gradient statistics, characterizing the dynamic characteristics of the local training process.

[0040] Homomorphic encryption is used to encrypt gradient statistics, generating encrypted statistics. Homomorphic encryption can support mathematical operations while the data is encrypted, ensuring that the encrypted gradient statistics will not leak the details of the original data when uploaded to the cloud, while allowing the cloud to directly perform calculations and analysis on the encrypted data.

[0041] By extracting the gradient mean and the magnitude of change, the dynamic characteristics of the local training process can be comprehensively described, providing valuable information for further analysis in the cloud. The use of homomorphic encryption technology not only protects the privacy and security of local training data, but also ensures the cloud's computing power for encrypted data, thereby improving the system's security and collaboration efficiency.

[0042] Step 203: Data Packaging and Uploading The initial weight vector generated after training and the encrypted gradient statistics are integrated into a complete data packet. The initial weight vector represents the preliminary weight distribution of local mining area monitoring indicators, while the encrypted gradient statistics provide a dynamic feature description of the local training process. This data packet is transmitted to the cloud aggregation service via a dedicated communication link to ensure the stability and security of the transmission process.

[0043] The uploaded data packet will serve as input for step three, used for cross-mining area data aggregation and the construction of potential weight manifolds; By integrating and packaging the initial weight vector and encrypted gradient statistics, locally generated data can be efficiently organized, facilitating transmission and cloud processing. Data is transmitted through a dedicated communication link, ensuring the reliability of the data during the upload process and providing the necessary basic information for cross-mine collaboration in the cloud.

[0044] Step two involves building and training a lightweight adaptive weighted network on local equipment in the mining area to generate an initial weight vector reflecting local risk characteristics. This initial weight vector, combined with encrypted gradient statistics, is then uploaded to the cloud, achieving efficient processing and secure sharing of local data. The entire process fully utilizes multi-source sensor data to generate a weight distribution adapted to the specific needs of each mining area. This overcomes the limitations of traditional fixed-weight models that cannot adapt to the differences between mining areas, demonstrating technological advantages in local data processing and privacy protection, and providing reliable support for cross-mining area collaboration.

[0045] Step 3: The cloud aggregation service receives encrypted update data from multiple mining areas, including encrypted gradient statistics and initial weight vectors. It decrypts the encrypted update data using a preset decryption key, generating decrypted gradient statistics and initial weight vectors. Then, it applies manifold learning techniques to the initial weight vectors, generating a latent weight manifold through dimensionality reduction and clustering to characterize the similarity and differences in weight distribution between mining areas. Simultaneously, it processes geological structure data, including geological age, fault development degree, and burial depth, generating geological feature vectors. An autoencoder is then used to convert these geological feature vectors into low-dimensional geological structure embedding vectors. Finally, it calculates the cosine similarity between the geological structure embedding vectors and the corresponding mining area coordinates in the latent weight manifold, quantifying the impact of geological features on weight distribution.

[0046] Step 301: Receive encrypted update data from multiple mining areas The cloud aggregation service first receives data packets uploaded from various mining nodes. These data packets contain encrypted gradient statistics and initial weight vectors. The gradient statistics are statistical data calculated by the mining nodes based on the local normalized feature tensors during local training of the lightweight adaptive weighted network, reflecting the gradient changes during local training. The initial weight vector is the preliminary weight distribution generated by the mining nodes through local training, representing the importance of each feature in the local model. The data is encrypted using homomorphic encryption before uploading to ensure data privacy and security during transmission.

[0047] The cloud aggregation service uses a pre-defined decryption key to decrypt the received encrypted data, restoring the encrypted gradient statistics and initial weight vectors into a data format suitable for subsequent processing. The decrypted gradient statistics and initial weight vectors will then serve as the foundational data for constructing potential patterns of weight distribution across mining areas, used for subsequent analysis and processing.

[0048] Homomorphic encryption effectively protects the privacy of local training data within the mining area, preventing sensitive information from being acquired or tampered with during transmission and ensuring data security. The decrypted gradient statistics and initial weight vectors provide diverse inputs from multiple mining areas for cloud-based aggregation services, enabling unified and consistent analysis of weight distribution across mining areas and thus improving the overall accuracy of risk assessment.

[0049] Step 302: Obtain the latent weighted manifold The cloud-based aggregation service utilizes manifold learning techniques to perform dimensionality reduction and clustering on the initial weight vectors from various mining areas to generate a latent weight manifold. The latent weight manifold is a low-dimensional space used to represent the similarities and differences between initial weight vectors from different mining areas, helping to analyze the intrinsic structure of the weight distribution.

[0050] The specific processing procedure is as follows: The cloud aggregation service treats the initial weight vector set of all mining areas as a set of high-dimensional data points. Through manifold learning algorithms (such as t-SNE or UMAP), these high-dimensional data points are projected into a low-dimensional space to form a latent weight manifold. This projection process preserves the local geometric relationships between the initial weight vectors. For example, some mining areas are close to each other due to similar weight distributions.

[0051] On the generated latent weight manifold, the cloud aggregation service applies clustering algorithms (such as K-means) to group low-dimensional data points, identifying similar and dissimilar patterns in the weight distribution. For example, some mining areas are grouped together due to similar feature importance distributions, while others belong to different groups due to significant differences. Manifold learning, through dimensionality reduction, clearly reveals the high-dimensional relationships between initial weight vectors, facilitating the comparison and understanding of weight distributions across mining areas. Cluster analysis further clarifies the common and unique characteristics of weight distributions across mining areas, providing a reliable basis for subsequent weight adjustment and mapping, and improving the efficiency and accuracy of cross-mining area analysis.

[0052] Step 303: Calculate the geological structure embedding vector The cloud aggregation service generates geological structure embedding vectors based on the geological structure data of each mining area. The geological structure data includes characteristics such as geological age, degree of fracture development, and burial depth. This data can be extracted from data packets uploaded by mining area nodes or retrieved from existing records in the cloud database.

[0053] First, the cloud aggregation service preprocesses the geological structure data to generate a geological feature vector. Specifically, it converts geological ages into discrete numerical representations, for example, older geological ages are represented by lower values ​​and newer geological ages by higher values; it normalizes continuous features such as fracture development and burial depth to a numerical range between 0 and 1; in this way, the geological structure data for each mining area is organized into a multidimensional geological feature vector.

[0054] Next, the cloud aggregation service uses an autoencoder to train the geological feature vectors and generate low-dimensional geological structure embedding vectors. The autoencoder learns the compressed representation of the input data, extracts the intrinsic connections and differences between geological features, and forms a low-dimensional vector that can reflect the geological structure features.

[0055] Finally, the cloud-based aggregation service calculates the correlation between the geological structure embedding vector and the corresponding mining area coordinates in the potential weight manifold.

[0056] The specific method involves comparing the similarity between the geological structure embedding vector and the mining area location in the potential weight manifold, using cosine similarity as the metric. Cosine similarity reflects the degree of influence of geological structure features on the weight distribution, providing reference information for subsequent weight adjustment.

[0057] By generating geological structure embedding vectors, the geological characteristics of the mining area are transformed into a low-dimensional form, facilitating correlation analysis with the potential weight manifold. The cosine similarity calculation method quantifies the relationship between geological structure and weight distribution, enabling weight adjustment to fully consider the impact of geological differences, thereby improving the adaptability and accuracy of the risk assessment method.

[0058] Step three involves receiving and decrypting encrypted update data from multiple mining areas, generating a latent weight manifold using manifold learning techniques, and calculating a geological structure embedding vector by combining this with geological structure data. This achieves preliminary integration of weight distribution across mining areas and quantitative representation of geological features. The latent weight manifold reveals the inherent patterns of weight distribution in each mining area, while the geological structure embedding vector incorporates the geological characteristics of the mining area into the weight analysis process. This design overcomes the mismatch problem caused by neglecting cross-mining area differences in traditional methods. Through the technical advantages of data aggregation and feature extraction, it provides important support for achieving dynamic and personalized risk assessment.

[0059] Step 4: The cloud aggregation service uses the geological structure embedding vector to map the initial weight vectors of each mining area to the potential weight manifold, and achieves alignment of the weight space by calculating a personalized transformation matrix. Specifically, the personalized transformation matrix is ​​generated by minimizing the squared Euclidean distance between the mapped weight vector and the mining area coordinate points in the potential weight manifold, and is fine-tuned and optimized based on the cosine similarity between the geological structure embedding vector and the mining area coordinate points. The optimized personalized transformation matrix is ​​sent back to the mining area nodes, and the mining area nodes use this personalized transformation matrix to adjust their local weight vectors for subsequent calculation of comprehensive disaster values.

[0060] Step 401, Weight Vector Mapping The cloud-based aggregation service first uses geological structure embedding vectors as the localization basis to map the initial weight vectors of each mining area onto a potential weight manifold. The geological structure embedding vectors, derived from the generation process in step three, reflect the geological characteristics of the mining area; the potential weight manifold represents the similarity and difference structure of the weight distribution; and the initial weight vectors are generated through local training in the mining area during step two, reflecting the initial contribution of local sensor data to disaster risk. The mapping process aims to adjust the initial weight vectors to align them with the coordinate points of the corresponding mining area in the potential weight manifold. The coordinate points in the potential weight manifold reflect the specific location of the mining area in terms of weight distribution characteristics. The mapping is achieved by calculating the transformation relationship between the initial weight vectors and the corresponding coordinate points.

[0061] A personalized transformation matrix is ​​generated for the i-th mining area. This matrix is ​​used to adjust the initial weight vector to its optimal position on the potential weight manifold. The calculation method involves selecting a transformation matrix that minimizes the distance between the mapped weight vector and the coordinate point. The distance is measured by summing the squares of the differences between the mapped weight vector and the coordinate point in Euclidean space.

[0062] Through optimization, a unique personalized transformation matrix is ​​determined. This transformation matrix is ​​applied to the initial weight vector to generate a mapped weight vector. The mapped weight vector approximates the coordinates of the corresponding mining area on the latent weight manifold, achieving alignment between the weight space and the unified cloud-based evaluation platform.

[0063] Through the mapping process, the weight vectors of each mining area are aligned with the potential weight manifold in the cloud, ensuring the consistency of weight distribution in cross-mining area assessments. The personalized transformation matrix is ​​customized based on the geological characteristics of each mining area, fully considering the differences between them and improving the targeting and accuracy of disaster risk assessment. This method utilizes geological characteristics to optimize weight distribution, avoiding the limitations of simple averaging or uniform processing.

[0064] Step 402: Generation and return of personalized transformation matrix To further improve mapping accuracy, the cloud-based aggregation service introduces geological structure embedding vectors to fine-tune the personalized transformation matrix. This fine-tuning is based on the similarity between the geological structure embedding vectors and the mining area coordinates in the potential weight manifold. The similarity is measured by calculating their cosine similarity; a higher cosine similarity indicates stronger consistency between the geological structure and the weight distribution. The fine-tuning mechanism adjusts the transformation matrix according to the similarity: when the cosine similarity is close to the maximum value of 1, it indicates a high degree of consistency between the geological structure and the weight distribution, requiring only minor adjustments to the transformation matrix; when the cosine similarity is low, the transformation matrix is ​​optimized using gradient descent to make the mapped weight vectors closer to the coordinate points in the potential weight manifold. The optimization objective is to minimize the sum of squared distances between the mapped weight vectors and the coordinate points, using cosine similarity as a constraint for iterative adjustments.

[0065] The optimized personalized transformation matrix is ​​transmitted to the corresponding mining area node via a dedicated communication link. Upon receiving the personalized transformation matrix, the mining area node applies it to local weight adjustments, generating an updated weight vector. This updated weight vector is used for subsequent calculations of the comprehensive disaster value, ensuring consistency between local assessments and cloud-based analysis.

[0066] Through a fine-tuning mechanism, the personalized transformation matrix can be optimized based on the geological characteristics of the mining area, ensuring the accuracy of the weight mapping. The use of cosine similarity quantifies the correlation between geological structure and weight distribution, enhancing the scientific rigor and adaptability of the assessment method. The personalized transformation matrix returned to the mining area nodes provides a precise adjustment basis for local risk assessment, improving the real-time performance and practicality of the assessment. This approach fully leverages cloud computing capabilities and local data characteristics, achieving efficient collaboration.

[0067] Step four utilizes cloud-based aggregation services to map the initial weight vectors of each mining area to the potential weight manifold using geological structure embedding vectors. A personalized transformation matrix is ​​then generated and sent back to the mining area nodes, achieving cross-domain alignment and personalized optimization of the weight space. This process effectively integrates the potential weight manifold and the geological structure embedding vectors, ensuring that the mapped weight vectors reflect both the geological differences of the mining areas and maintain consistency with the cloud-based assessment platform. Through the technical advantages of data mapping and feature fusion, a reliable weight adjustment mechanism is provided for comprehensive disaster assessment, improving the accuracy and applicability of the assessment method. This method demonstrates significant flexibility and scientific rigor in cross-mining area risk assessment.

[0068] Step 5: After receiving the personalized transformation matrix, the mining area node calculates the comprehensive disaster value using the personalized transformation matrix and the standardized feature tensor. Then, it generates a real-time scheduling priority coefficient by jointly quantizing the risk growth signal and the response lag signal. Based on the real-time scheduling priority coefficient, the control commands are sorted and issued, and the performance feedback indicators are uploaded to the cloud. Step 501: Calculation of Comprehensive Disaster Values The mining node first adjusts the initial weight vector using a personalized transformation matrix transmitted from the cloud, generating an updated weight vector. This personalized transformation matrix, generated by the cloud aggregation service in step four, aligns the local weight vector with the potential weight manifold. The initial weight vector originates from the training results of the mining area's local adaptive weighted network in step two. The updated weight vector is calculated by multiplying the personalized transformation matrix by the initial weight vector. This process multiplies each component of the initial weight vector with the corresponding row element of the personalized transformation matrix and sums them, reflecting the optimized distribution of local weights after cross-domain alignment. Next, the mining node uses the updated weight vector and a standardized feature tensor for weighted summation to calculate the comprehensive hazard value. The standardized feature tensor, generated in step one, contains sensor data that has undergone synchronous calibration and confidence screening.

[0069] The weighted summation process involves multiplying each component of the standardized feature tensor with the corresponding component of the updated weight vector and then summing them all to generate a scalar value, namely the comprehensive disaster value, which represents the overall disaster risk level of the current mining area.

[0070] The personalized transformation matrix ensures consistency between local weights and the cloud-based assessment platform while preserving the unique impact of the mining area's geological characteristics. The weighted summation method is computationally simple and can quickly quantify disaster risks. This design guarantees assessment accuracy while meeting the stringent computational speed requirements of the underground environment, providing immediate data for real-time scheduling.

[0071] Step 502, Calculation of Risk Growth Signal By calculating the deviation between the comprehensive disaster value and its historical trend, and combining it with sensor confidence levels, the mining area nodes generate a risk growth signal.

[0072] First, an exponentially weighted moving average of the comprehensive disaster value is calculated to smooth historical data and reflect trends. The calculation process involves weighting the current comprehensive disaster value with the previous exponentially weighted moving average. The current comprehensive disaster value is multiplied by a smoothing coefficient, and the previous exponentially weighted moving average is multiplied by the difference between the current and previous values ​​minus the smoothing coefficient. These two values ​​are then added together to obtain the new exponentially weighted moving average. The smoothing coefficient controls the weighting ratio between the old and new data. Next, the risk growth signal is calculated using the following steps: the current comprehensive disaster value is subtracted from the newly obtained exponentially weighted moving average, and the difference is multiplied by the square root of the sensor confidence level. The sensor confidence level, derived from step one, reflects data reliability. Its square root is calculated using the positive square root of the confidence level and is used for weighted adjustment, giving higher-confidence data a greater impact from the risk growth rate. The risk growth signal is a dimensionless scalar, quantifying the instantaneous speed of risk growth.

[0073] Exponentially weighted moving averages smooth historical trends and avoid interference from short-term fluctuations; square root adjustment of sensor confidence enhances the impact of data quality on assessment. This approach provides a real-time and reliable indicator of risk growth rate, supporting the rapid identification of potential disaster trends.

[0074] Step 503: Calculation of response hysteresis signal By analyzing the correlation between recent regulatory actions and changes in comprehensive disaster magnitude, the mining area node generates a response lag signal: First, the cross-correlation coefficient between the intensity of the near-term control action and the comprehensive disaster value after the lag time is calculated to identify the degree of impact of the control action on risk changes. The calculation process of the cross-correlation coefficient is to normalize the covariance of the control action sequence and the lagged comprehensive disaster value sequence, that is, to divide the covariance of the two sequences by the product of their respective standard deviations. The lag time is tried one by one within a preset range, and the value that maximizes the cross-correlation coefficient is selected.

[0075] Subsequently, the response lag signal is calculated by dividing the cross-correlation coefficient by the historical system inertia constant. The historical system inertia constant, based on the statistical analysis of response times of historical control actions and risk changes, reflects the inertial characteristics of the mining area system, and its value is determined through historical data analysis. The response lag signal is a dimensionless scalar, quantifying the delay in the effectiveness of control actions. Cross-correlation analysis accurately captures the temporal relationship between control actions and risk changes, and the introduction of the historical system inertia constant adapts the signal to the characteristics of different mining areas. This method provides a quantitative perspective on the delay in the effectiveness of control actions, enhances the scientific nature of scheduling decisions, and avoids blind intervention.

[0076] Step 504: Real-time scheduling priority coefficient generation Mining nodes input risk growth signals and response lag signals into a pre-trained interpretable decision surface model (IDFSM) to generate real-time scheduling priority coefficients. The IDFSM, based on support vector machines or decision trees and trained using historical data, outputs scheduling priorities based on the input signals. The model's inputs are the risk growth signal and the response lag signal, and its output is a dimensionless scalar representing the scheduling priority. The model training process uses historical risk growth signals, response lag signals, and corresponding actual scheduling priority data, adjusting model parameters to minimize the difference between predicted and actual values. After training, the model calculates real-time scheduling priority coefficients by comparing the current risk growth signal and response lag signal with internal decision rules. The IDFSM, combined with quantitative indicators of risk growth and control delay, ensures a transparent and traceable decision-making process; the pre-training method improves computational efficiency. This design balances accuracy and timeliness, and the generated scheduling priority coefficients directly support the scientific sequencing of on-site control commands, adapting to the real-time needs of underground scheduling.

[0077] Step 505: Control Command Issuance and Performance Feedback Mining area nodes prioritize control commands based on real-time scheduling priority coefficients, executing commands from areas with higher priority coefficients first. The prioritization process involves arranging all areas by real-time scheduling priority coefficients from highest to lowest, generating a control command queue accordingly. After a command is issued, the execution effect and execution time data are recorded. The execution effect includes changes in the overall disaster magnitude, and the execution time is the time interval from command issuance to its effective date.

[0078] Subsequently, the mining nodes upload execution results and risk change data to the cloud as performance feedback indicators. These performance feedback indicators include quantitative information such as changes in the overall disaster magnitude and response time, used for updating the potential weight manifold in step six. Scheduling priority ranking ensures that high-risk areas receive priority intervention; the uploading of performance feedback indicators provides real-time data support for weight adjustments in the cloud. This mechanism improves the targeting and efficiency of scheduling, while enhancing the system's dynamic adaptability through closed-loop feedback, ensuring long-term operational stability.

[0079] By calculating comprehensive disaster values ​​using personalized transformation matrices and standardized feature tensors at mining area nodes, and combining these with risk growth signals and response lag signals to generate real-time scheduling priority coefficients, real-time quantification and scheduling optimization of coal and rock dynamic disaster risks are achieved. The calculation process is rigorous and efficient, and the technical features are scientifically sound, ensuring the accuracy and practicality of the assessment results.

[0080] The combined application of risk acceleration signals and response lag signals provides a multi-dimensional perspective for risk assessment. The generation of real-time scheduling priority coefficients directly supports the scientific sequencing of on-site control commands, meeting the timeliness requirements of downhole scheduling. The synchronous uploading of performance feedback indicators provides a data foundation for continuous system optimization, demonstrating the technical advantages of this invention in dynamic risk assessment and collaborative scheduling.

[0081] Step Six: The technical features of the cloud aggregation service periodically updating the potential weight manifold and generating personalized transformation matrices are as follows: The cloud aggregation service first collects performance feedback indicators and geological drift signs from each mining area node. Then, it constructs a performance feedback matrix based on the performance feedback indicators and a drift sign vector based on the geological drift signs. It calculates the comprehensive performance score and influence factor through the performance feedback matrix and the drift sign vector. After that, it updates the potential weight manifold using the comprehensive performance score and influence factor. Based on the updated potential weight manifold, it generates a personalized transformation matrix for each mining area node. Finally, it transmits the personalized transformation matrix to the corresponding mining area node through a dedicated communication link.

[0082] Step 601: Collection of cumulative performance feedback indicators and geological drift signs The cloud aggregation service first receives performance feedback metrics and geological drift indicators from each mining node.

[0083] Performance feedback metrics are derived from the data uploaded by the mining nodes in step five, reflecting the execution effectiveness of scheduling instructions and risk control performance. These metrics include quantitative information such as changes in comprehensive disaster magnitude and response time. The cloud aggregation service summarizes the performance feedback metrics of all mining nodes according to a preset time window, such as weekly or monthly, generating a performance feedback matrix. The number of rows in the performance feedback matrix equals the number of mining areas, and the number of columns equals the number of performance metrics. Geological drift indicators are obtained by analyzing the changing trends of geological structure data in the mining area, involving changes in characteristics such as the degree of fracture development and burial depth. The analysis method employs a sliding window variation point detection technique to process the time series of geological structure data, identify significant change points, and generate a drift indicator vector. The dimension of the drift indicator vector is the same as the number of geological features. This performance feedback index quantifies the system's performance in actual operation, while geological drift indicators capture the dynamic changes in the geological conditions of the mining area. The combination of these two provides comprehensive data support for updating the potential weight manifold. This data collection method ensures that the update process considers both system performance and environmental changes, thereby improving the adaptability of the evaluation method.

[0084] Step 602: Update of the latent weighted manifold The cloud aggregation service updates the underlying weight manifold using the performance feedback matrix and drift indication vector.

[0085] First, a comprehensive performance score is calculated for each mining area based on the performance feedback matrix. To calculate the comprehensive performance score, the performance index value of each mining area in the performance feedback matrix is ​​multiplied by a preset performance index weight. Then, all weighted results are summed to obtain a scalar value representing the overall performance of that mining area. The performance index weights are determined through analysis of historical data to ensure that the score accurately reflects the system's operational status.

[0086] Secondly, the influence factor of geological changes on the potential weight manifold is calculated based on the drift sign vector. When calculating the influence factor, the drift sign vector is multiplied by the pre-trained influence matrix. The influence matrix is ​​obtained by learning the relationship between historical geological data and weight changes, reflecting the impact of geological feature changes on the weight distribution. The dimension of the influence factor is consistent with the dimension of the potential weight manifold. Subsequently, the potential weight manifold is updated using the comprehensive performance score and the influence factor. The update method is to add the original potential weight manifold with a weighted sum of the comprehensive performance score and the influence factor, where the weighting coefficients are controlled by the learning rate. The learning rate is a preset constant used to adjust the update step size.

[0087] Using this comprehensive performance score as the basis for weight adjustment ensures that the update direction is consistent with the system performance; the introduction of influencing factors allows the update process to adapt to changes in geological conditions; and the learning rate balances the stability and responsiveness of the update. This update mechanism enables the potential weight manifold to dynamically reflect the actual situation in the mining area, thereby improving the long-term stability of the evaluation method.

[0088] Step 603: Regenerate the transformation matrix and synchronize it to the mining node. The cloud aggregation service regenerates a personalized transformation matrix for each mining area based on the updated potential weight manifold. The generation method is the same as in step four, and the new matrix is ​​calculated by optimizing the mapping error. Specifically, the process of generating the personalized transformation matrix aims to minimize the squared Euclidean distance between the transformed initial weight vector of the mining area and the coordinates of the mining area in the updated potential weight manifold. The initial weight vector is derived from step two, and the coordinates of the mining area are the positions of the corresponding mining areas in the updated potential weight manifold.

[0089] The optimization process ensures that the new transformation matrix accurately maps the initial weight vectors to the updated potential weight manifold. After generating the new transformation matrix, the cloud aggregation service transmits it to the corresponding mining area nodes via a dedicated communication link. Upon receiving the new transformation matrix, each mining area node updates its local weight adjustment mechanism for subsequent calculations of comprehensive hazard values ​​and real-time scheduling priority coefficients. This regenerated transformation matrix reflects the latest weight distribution pattern, ensuring that mining area nodes can adjust weights according to current geological conditions and system performance. This synchronization mechanism enables the assessment method to adapt to changes in real time, improving dynamic response capabilities and assessment accuracy.

[0090] Step six utilizes cloud aggregation services to periodically update the potential weight manifold based on accumulated performance feedback metrics and geological drift indicators, generating new personalized transformation matrices and synchronizing them to the mining area nodes. This enables dynamic adjustment and continuous optimization of the evaluation method. The entire process comprehensively considers the impact of system performance and geological changes, ensuring that the potential weight manifold accurately reflects the actual conditions of the mining area. The generation and synchronization of the new transformation matrix provides the mining area nodes with the latest weight adjustment basis, improving the adaptability of the evaluation method to environmental changes and enhancing the long-term stability of the system.

[0091] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0092] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0093] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0094] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0095] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An adaptive variable weighted comprehensive assessment method for coal and rock dynamic hazards, characterized in that: include, Each mining area performs synchronous calibration and reliability screening of underground multi-source sensor streams locally, generates standardized feature tensors and calculates the corresponding encrypted gradient statistics; Mining nodes train a lightweight adaptive weighted network using standardized feature tensors as input, obtain an initial weight vector, and package it together with encrypted gradient statistics to the cloud. The cloud aggregation service receives encrypted updated data from multiple mining areas, applies manifold learning techniques to the initial weight vectors, and generates a latent weight manifold through dimensionality reduction and clustering to characterize the similarity and differences in weight distribution among mining areas. Simultaneously, it processes geological structure data, including geological age, fault development degree, and burial depth, to generate geological feature vectors. An autoencoder is then used to convert these geological feature vectors into low-dimensional geological structure embedding vectors. The cosine similarity between the geological structure embedding vectors and the corresponding mining area coordinates in the latent weight manifold is calculated to quantify the impact of geological features on weight distribution. Specifically, the cloud aggregation service treats the initial weight vector set of all mining areas as a set of high-dimensional data points and projects these high-dimensional data points into a low-dimensional space using a manifold learning algorithm to form a latent weight manifold. This projection process preserves the local geometric relationships between the initial weight vectors. The cloud aggregation service uses geological structure embedding vectors to map the initial weight vectors of each mining area to the potential weight manifold, and achieves alignment of the weight space by calculating a personalized transformation matrix. The personalized transformation matrix is ​​generated by minimizing the squared Euclidean distance between the mapped weight vector and the mining area coordinates in the potential weight manifold, and is fine-tuned and optimized based on the cosine similarity between the geological structure embedding vector and the mining area coordinates. The optimized personalized transformation matrix is ​​then sent back to the mining area nodes. After receiving the transformation matrix, the mining area node immediately calculates the comprehensive disaster value and generates a real-time scheduling priority coefficient by jointly quantifying the risk growth signal and the response lag signal. Based on this, control commands are issued and performance feedback indicators are synchronized to the cloud. Based on accumulated performance feedback metrics and geological drift indicators, the cloud periodically updates the potential weight manifold, regenerates the transformation matrix, and synchronizes it to all mining area nodes.

2. The adaptive variable weight coal and rock dynamic disaster comprehensive assessment method as described in claim 1, characterized in that: The underground multi-source sensor streams are processed locally in the mining area. The multi-source sensor data includes: earthquake data, surrounding rock stress data, gas content data, and temperature and humidity data. The raw calibration dataset is generated through time alignment and measurement calibration. Based on historical data and sensor performance characteristics, sliding window analysis and isolated forest algorithms are used to identify anomalous data points and remove unreliable data to generate a reliable dataset.

3. The adaptive variable weight coal and rock dynamic disaster comprehensive assessment method as described in claim 2, characterized in that: Temporal, frequency, and spatial features are extracted from a trusted dataset, and standardized to eliminate dimensional differences, forming a standardized feature tensor. Homomorphic encryption is then used to encrypt the standardized feature tensor, generating an encrypted feature tensor. Encrypted gradient statistics are then calculated based on the encrypted feature tensor.

4. The adaptive variable weight coal and rock dynamic disaster comprehensive assessment method as described in claim 3, characterized in that: During the training of the lightweight adaptive weighted network, the nodes in the mining area record the gradient changes of the network parameters, calculate the mean gradient and the magnitude of gradient changes to generate gradient statistics, and encrypt the gradient statistics to form encrypted gradient statistics.

5. The adaptive variable weight coal and rock dynamic disaster comprehensive assessment method as described in claim 4, characterized in that: Receive encrypted update data from multiple mining areas, including encrypted gradient statistics and initial weight vectors, decrypt the encrypted update data using a preset decryption key, and generate decrypted gradient statistics and initial weight vectors.

6. The adaptive variable weight coal and rock dynamic disaster comprehensive assessment method as described in claim 5, characterized in that: After receiving the personalized transformation matrix, the mining node calculates the comprehensive disaster value using the personalized transformation matrix and the standardized feature tensor. It generates a real-time scheduling priority coefficient by jointly quantizing the risk growth signal and the response lag signal, and sorts and issues control commands according to the real-time scheduling priority coefficient. At the same time, it uploads the performance feedback indicators.

7. The adaptive variable weight coal and rock dynamic disaster comprehensive assessment method as described in claim 6, characterized in that: Collect performance feedback indicators and geological drift signs of each mining area node, construct a performance feedback matrix based on the performance feedback indicators, and construct a drift sign vector based on the geological drift signs; The comprehensive performance score and influence factor are calculated using the performance feedback matrix and drift indicator vector. The potential weight manifold is updated using the comprehensive performance score and influence factor. Based on the updated potential weight manifold, a personalized transformation matrix is ​​generated for each mining node. The personalized transformation matrix is ​​transmitted to the corresponding mining node through a dedicated communication link.

Images