Ultra-deep underground space remote monitoring method and system based on multi-source data fusion

Abstract

The invention discloses an ultra-deep underground space remote monitoring method and system based on multi-source data fusion, and the method comprises the steps: obtaining initial data from a multi-source device through a preset interface, and unifying the initial data into a sequence containing time and a numerical value through a conversion module; grouping the embedding positions of the similar equipment by using a clustering algorithm, and generating a labeled integrated data set; associating environment variables, and if the coefficient exceeds a threshold value, fusing into a dynamic monitoring matrix; extracting trends from the data, identifying abnormalities through a classification model, and outputting an annotation mapping layer; generating transmission protocol data according to the transmission protocol data, and optimizing the transmission protocol data into a signal sequence integrating a variable relationship through a computational node; and the final prediction model estimates the trend key points and generates a risk distribution diagram corresponding to the monitoring state. According to the method, a fusion mechanism of a matrix and a signal sequence is dynamically monitored, and closed-loop circulation of data from heterogeneity to optimization is realized, so that high-risk distribution is accurately predicted, and the environment monitoring response efficiency is improved; a real-time and accurate risk early warning system is constructed, and the loss of emergencies is remarkably reduced.

Description

Technical Field

[0001] This invention relates to the field of ultra-deep underground space monitoring technology, and in particular discloses a method and system for remote monitoring of ultra-deep underground space based on multi-source data fusion. Background Technology

[0002] Monitoring of ultra-deep underground spaces is a crucial area for ensuring the safety of underground engineering projects and the stability of urban construction. It involves the real-time acquisition and analysis of multi-dimensional data, including groundwater levels, structural deformation, and stress changes, directly impacting engineering safety and disaster prevention. Precise monitoring not only reduces engineering accidents but also provides critical data for urban planning and resource management. However, existing monitoring methods often face challenges in effectively integrating and analyzing multi-source data in real-time when dealing with complex underground environments, limiting the adaptability and early warning capabilities of monitoring systems.

[0003] The current limitations of monitoring technology lie in the significant differences in data acquisition and processing methods among various sensors, leading to difficulties in data integration. For example, voltage-current type, vibrating wire type, and inductive type sensors differ significantly in their working principles, data formats, and acquisition frequencies, making it difficult for existing equipment to uniformly process this heterogeneous data. Furthermore, the correlation between geospatial information and sensor data is insufficient, making it difficult to achieve an effective correspondence between spatial location and monitoring data. This results in low efficiency in data acquisition and analysis in the complex environment of ultra-deep underground spaces, affecting the comprehensiveness and accuracy of monitoring.

[0004] The core technical challenges lie in the integrated acquisition of multi-source heterogeneous data and the fusion of geospatial information with monitoring data. Inconsistent acquisition command protocols among different sensors prevent data acquisition equipment from efficiently supporting multiple sensors. For example, vibrating wire sensors have a high acquisition frequency but small data volume, while inductive sensors generate large data volumes but have slow processing speeds. This difference complicates the design of a unified acquisition protocol. Furthermore, the lack of an efficient mapping mechanism between geospatial information and sensor data prevents monitoring data from accurately reflecting the actual state of underground space. For instance, when monitoring groundwater level changes, sensor data may not be correlated with specific geographical locations, making it difficult to determine the impact of water level changes on the structure of a particular area.

[0005] Therefore, designing a unified acquisition protocol and data fusion mechanism to achieve efficient acquisition, transmission, and accurate mapping of multi-source heterogeneous sensor data with geospatial information has become a key issue in ultra-deep underground space monitoring. Summary of the Invention

[0006] This invention provides a method and system for remote monitoring of ultra-deep underground space based on multi-source data fusion, aiming to solve at least one of the defects existing in the above-mentioned prior art.

[0007] One aspect of the present invention relates to a remote monitoring method for ultra-deep underground space based on multi-source data fusion, comprising the following steps:

[0008] S100: Obtain initial data from multi-source monitoring devices through preset interface specifications, use conversion module to process different format differences, and obtain a unified data sequence. The unified data sequence contains time information and numerical markers.

[0009] S200. Based on the unified data sequence, a clustering algorithm is used to group similar device types, determine the group subsets, and embed location information in the group subsets to obtain an integrated dataset with annotations.

[0010] S300. For the integrated dataset, obtain the correlation information of environmental variable changes. If the coefficient between the correlation information exceeds the threshold, it is fused into a dynamic matrix to obtain a monitoring matrix that reflects the distribution.

[0011] S400. Extract trend information from the monitoring matrix, use a classification model to identify abnormal patterns, determine high-risk areas, and obtain a mapping layer that marks anomalies.

[0012] S500: Based on the mapping layer, generate transmission protocol data, use computing nodes to process the information flow within the transmission protocol data, obtain an optimized signal sequence, and integrate the variable correspondence of the signal sequence;

[0013] S600. For the signal sequence, a prediction model is used to estimate the trend information, determine the location of key points in the trend information, and obtain a risk distribution map, which corresponds to the monitoring status.

[0014] Further, step S100 includes:

[0015] S110. Obtain initial data from multi-source monitoring equipment, and standardize and parse the initial data using a preset interface specification to obtain an initial data set containing timestamps and values.

[0016] S120. To address the format differences in the initial dataset, a conversion module is used to process the standardized parsed initial data into a unified format, generating a unified data sequence containing time information and numerical markers.

[0017] Further, step S200 includes:

[0018] S210. Based on the unified data sequence, the K-means clustering algorithm is used to group the devices by similarity, generating a subset of groups containing similar device types;

[0019] S220. Based on the group subsets, obtain the location information of each group of devices, and combine the location information with the group subsets through the location embedding module to generate an intermediate dataset with location annotations.

[0020] S230. If the location labels in the intermediate dataset are complete, the time information, numerical labels and location labels are merged through the data integration module to obtain an integrated dataset with labels.

[0021] Further, step S300 includes:

[0022] S310. Based on the integrated dataset, obtain environmental variable change data from environmental variables, calculate the Pearson correlation coefficient between environmental variable changes and equipment status through the correlation analysis module, and obtain the correlation coefficient matrix. The environmental variable change data includes temperature change data, humidity change data, and light intensity change data.

[0023] S320. If at least one coefficient in the correlation coefficient matrix exceeds a preset threshold, the environmental variable change data and equipment status data are weighted and combined through the matrix fusion module to generate a dynamic matrix.

[0024] S330. Principal component analysis algorithm is used to reduce the dimensionality of the dynamic matrix, extract the main distribution features, and obtain the dimensionality-reduced feature matrix.

[0025] S340. The data points of the feature matrix are mapped to the preset monitoring grid through the distribution mapping module to generate a monitoring matrix that reflects the distribution relationship between environmental variables and equipment status.

[0026] Further, step S400 includes:

[0027] S410. Based on the monitoring matrix, extract the trend sequence of environmental variables through a time series filter, calculate the difference value of consecutive time points in the trend sequence of environmental variables, and if the difference value exceeds the preset threshold, mark the potential change points to obtain the trend sequence of marked change points.

[0028] S420. Use a support vector machine to classify the trend sequence of labeled change points, distinguish between potential change points and normal points, and generate a classification label sequence.

[0029] S430. Obtain the dense region of abnormal labels from the classification label sequence, determine the boundary of the high-risk region through clustering algorithm, and if the abnormal density in the high-risk region exceeds the preset threshold, fuse the boundary information to obtain the outline of the high-risk region.

[0030] S440. Using a raster rendering tool, overlay the outline of the high-risk area onto the corresponding position of the monitoring matrix, mark the position of the abnormal mode, and generate a mapping layer that marks the abnormality.

[0031] Further, step S500 includes:

[0032] S510. Based on the mapping layer, extract the coordinate data of the high-risk area outline, and use a data encoding tool to serialize the coordinate data to obtain a protocol data packet containing area information.

[0033] S520. The information flow in the protocol data packet is fragmented by the computing node. If the number of fragments exceeds the preset threshold, the fragments are reassembled using a parallel processing tool to obtain the reassembled information flow sequence.

[0034] S530. Use a convolutional neural network to extract features from the recombined information flow sequence, obtain optimized signal features, and obtain a signal sequence containing the correspondence between variables.

[0035] Further, step S600 includes:

[0036] S610. For the signal sequence, a time series prediction model is used to estimate future trend information, and the peak point coordinates are extracted from the future trend information to obtain a preliminary trend curve.

[0037] S620. Based on the preliminary trend curve, a clustering algorithm is used to group the peak point locations. If the group density exceeds a preset threshold, noise points are removed using a density filtering tool to determine the key point location sequence.

[0038] S630. Based on the key point location sequence, the peak point locations are converted into a two-dimensional coordinate grid using a rasterization mapping tool. The risk intensity value of each grid is obtained from the two-dimensional coordinate grid to obtain the risk distribution matrix.

[0039] S640. For the risk distribution matrix, a state mapping table is used to convert the risk intensity value into a monitoring state label. If the monitoring state label meets the preset level, the state level of each area is determined to obtain the risk distribution map of the corresponding monitoring state.

[0040] Another aspect of the present invention relates to a remote monitoring system for ultra-deep underground space based on multi-source data fusion, used to execute the above-described remote monitoring method for ultra-deep underground space based on multi-source data fusion, comprising:

[0041] The unified data sequence acquisition module is used to acquire initial data from multi-source monitoring devices through a preset interface specification, and uses a conversion module to process the differences in different formats to obtain a unified data sequence. The unified data sequence contains time information and numerical markers.

[0042] The integrated dataset acquisition module is used to group similar device types according to a unified data sequence using a clustering algorithm, determine the group subsets, and embed location information in the group subsets to obtain an integrated dataset with annotations.

[0043] The monitoring matrix acquisition module is used to acquire correlation information of environmental variable changes for an integrated dataset. If the coefficients between the correlation information exceed a threshold, they are fused into a dynamic matrix to obtain a monitoring matrix that reflects the distribution.

[0044] The mapping layer acquisition module is used to extract trend information from the monitoring matrix, identify abnormal patterns using a classification model, determine high-risk areas, and obtain a mapping layer that marks anomalies.

[0045] The signal sequence acquisition module is used to generate transmission protocol data based on the mapping layer, process the information flow within the transmission protocol data using computing nodes, and obtain an optimized signal sequence. The signal sequence integrates the correspondence between variables.

[0046] The risk distribution map acquisition module is used to estimate trend information for signal sequences using a prediction model, determine the location of key points in the trend information, and obtain a risk distribution map, which corresponds to the monitoring status.

[0047] The beneficial effects achieved by this invention are as follows:

[0048] This invention discloses a method and system for remote monitoring of ultra-deep underground spaces based on multi-source data fusion. It addresses the operational challenges of data silos caused by format differences among multi-source devices, difficulties in grouping similar devices, and inaccurate fusion of environmental variable correlations. These problems converge to form a logical bottleneck preventing real-time mapping of high-risk areas. Initial data is acquired from multi-source devices via a pre-defined interface and converted into a time- and numerical sequence by a conversion module. A clustering algorithm then groups similar devices by their embedded locations, generating a labeled integrated dataset. Environmental variables are then correlated; if coefficients exceed thresholds, the data is fused into a dynamic monitoring matrix. Trends are extracted, and anomalies are identified using a classification model, outputting a labeled mapping layer. Transmission protocol data is generated and optimized by computation nodes into a signal sequence integrating variable relationships. Finally, a prediction model estimates key trend points, generating a risk distribution map corresponding to the monitoring status. The most important inventive aspect lies in the fusion mechanism of the dynamic monitoring matrix and signal sequence, achieving a closed-loop flow of data from heterogeneous to optimized data, thereby accurately predicting high-risk distributions and improving environmental monitoring response efficiency. The overall technical effect is to construct a real-time and accurate risk early warning system, significantly reducing losses from emergencies. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating an embodiment of the remote monitoring method for ultra-deep underground space based on multi-source data fusion according to the present invention.

[0050] Figure 2 This is a functional block diagram of an embodiment of the ultra-deep underground space remote monitoring system based on multi-source data fusion of the present invention.

[0051] Explanation of icon numbers:

[0052] 10. Unified data sequence acquisition module; 20. Integrated dataset acquisition module; 30. Monitoring matrix acquisition module; 40. Mapping layer acquisition module; 50. Signal sequence acquisition module; 60. Risk distribution map acquisition module. Detailed Implementation

[0053] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0054] like Figure 1 As shown, the first embodiment of the present invention proposes a remote monitoring method for ultra-deep underground space based on multi-source data fusion, comprising the following steps:

[0055] Step S100: Obtain initial data from multi-source monitoring devices through preset interface specifications, use conversion module to process different format differences, and obtain a unified data sequence. The unified data sequence contains time information and numerical markers.

[0056] Based on the pre-defined interface specifications (such as Modbus and OPC UA protocols) for ultra-deep underground space monitoring requirements, initial data is synchronously collected from multiple sources, including structural monitoring (fiber optic grating sensors), environmental monitoring (gas / temperature and humidity sensors), and equipment monitoring (vibration / current sensors). Addressing the format differences in output from different devices (such as binary, ASCII, and JSON), a conversion module performs format parsing, unit unification, and field mapping, ultimately generating a unified data sequence containing "timestamp (accurate to milliseconds) - monitoring value - device identifier." This provides standardized and correlated basic data input for subsequent data integration and fusion.

[0057] Step S200: Based on the unified data sequence, a clustering algorithm is used to group similar device types, determine the group subsets, and embed location information in the group subsets to obtain an integrated dataset with annotations.

[0058] Using the unified data sequence (including device ID, monitoring values, and time information) generated in step S100 as input, the core features of the devices (such as sampling frequency, monitoring dimensions, and value fluctuation range) are extracted. K-means or density clustering (DBSCAN) algorithms are used to group devices with similar functions and data characteristics into the same group, forming several group subsets (such as "structural displacement monitoring subset" and "environmental gas monitoring subset"). Then, the device ID is associated with a preset spatial location database (including sensor installation coordinates and the area to which it belongs) to embed the corresponding location information (such as the 300m underground dome and the 280m sidewall) into each group subset. Finally, an integrated dataset with annotations containing four-dimensional information of "device type group, monitoring data, timestamp, and spatial location" is generated, providing a data foundation with clear classification and spatial traceability for subsequent environmental variable correlation analysis.

[0059] Step S300: For the integrated dataset, obtain the correlation information of environmental variable changes. If the coefficients between the correlation information exceed the threshold, they are fused into a dynamic matrix to obtain a monitoring matrix that reflects the distribution.

[0060] Using the labeled integrated dataset (including equipment groups, monitoring data, and spatial locations) generated in step S200 as input, the system focuses on environmental monitoring-related subsets (such as CO2 concentration, temperature and humidity, and wind speed), extracts the change data of different environmental variables in the same time and adjacent spaces, and calculates the correlation coefficients between variables (such as Pearson correlation coefficient and mutual information value). If the correlation coefficient exceeds a preset threshold (such as |r|≥0.7), it is determined that the variables are significantly correlated. The time-series-spatial data of these correlated variables are fused into a dynamic matrix according to the "time slice-spatial grid" structure. The correlation strength and numerical distribution of variables in each spatial grid are quantified by matrix elements, and finally a monitoring matrix reflecting the correlation relationship and spatial distribution characteristics of environmental variables is obtained, providing correlated and spatialized data support for subsequent anomaly identification and risk location.

[0061] Step S400: Extract trend information from the monitoring matrix, use a classification model to identify abnormal patterns, determine high-risk areas, and obtain a mapping layer with anomalies marked.

[0062] Using the monitoring matrix (including spatiotemporal grids, values ​​of related variables, and correlation coefficients) generated in step S300 as input, the changing trend information of environmental variables within each spatial grid (such as the rate of increase / decrease in values ​​and the fluctuation range of correlation coefficients) is extracted through time series analysis. Using a preset classification model (such as random forest or support vector machine), the real-time trend information is compared with the historical normal pattern baseline to identify abnormal patterns that deviate from the normal range (such as sudden increases in variable values ​​or breaks in correlation relationships). High-risk areas are determined by combining the intensity of the anomaly (such as the number of times the value exceeds the threshold) and spatial clustering (such as anomalies in continuous grids). Finally, the anomaly pattern type, high-risk level, and spatial grid are associated to generate a two-dimensional / three-dimensional mapping layer that marks the location and risk level of the anomaly, providing an intuitive spatial anomaly basis for subsequent remote control and emergency response.

[0063] Step S500: Generate transmission protocol data based on the mapping layer, use computing nodes to process the information flow within the transmission protocol data, obtain an optimized signal sequence, and integrate the variable correspondence of the signal sequence.

[0064] Using the anomaly-marked mapping layer (including anomaly location, risk level, and anomaly type) generated in step S400 as input, the spatial grid information, anomaly feature data (such as numerical deviation and risk score), and associated variable identifiers in the layer are packaged into transmission protocol data according to the format specifications of a preset transmission protocol (such as MQTT or Modbus TCP). The protocol data is then processed by distributed computing nodes (such as data cleaning, redundancy removal, and priority sorting) to remove invalid fields (such as redundant historical data in low-risk areas) and retain core information (such as real-time anomaly sequences in high-risk areas). Finally, the processed information is integrated according to the logic of "timestamp-spatial grid-variable value-correlation relationship" to generate an optimized signal sequence. This optimized signal sequence clarifies the correspondence between various monitored variables (such as the anomaly correlation between CO2 concentration and temperature in a certain grid), providing a structured and timely signal input for subsequent trend prediction.

[0065] Step S600: For the signal sequence, use a prediction model to estimate the trend information, determine the location of key points in the trend information, and obtain a risk distribution map. The risk distribution map corresponds to the monitoring status.

[0066] Using the optimized signal sequence (including time-space-variable correlation information) generated in step S500 as input, a combination of time-series prediction models (such as LSTM, ARIMA) and spatial diffusion models (such as Gaussian diffusion, cellular automata) is employed to estimate the changing trend information of each monitored variable (such as the rate of increase in CO2 concentration and the probability of correlation recovery) within a preset time period (such as 10-60 minutes). By setting thresholds (such as values ​​exceeding the safety limit or trend abrupt change points), key time points (such as CO2 reaching the warning value after 15 minutes) and key spatial points (such as the risk diffusion source grid) are extracted from the trend information to determine the location of key points. Finally, the trend prediction results and key point locations are integrated and superimposed with the spatial base of ultra-deep underground space (such as BIM model) to generate a risk distribution map labeled with risk level, diffusion path, and key point coordinates. This risk distribution map corresponds one-to-one with the real-time monitoring status, providing an intuitive risk prediction basis for remote control and emergency response.

[0067] Furthermore, the ultra-deep underground space remote monitoring method based on multi-source data fusion proposed in this embodiment includes step S100 as follows:

[0068] Step S110: Obtain initial data from multi-source monitoring equipment, and standardize and parse the initial data using a preset interface specification to obtain an initial data set containing timestamps and values.

[0069] The initial dataset, containing timestamps and numerical values, is derived using the following formula:

[0070]

[0071] In formula (1), D s This represents the initial data set containing timestamps and values, where M represents the total number of monitoring devices, and P... i R represents the default interface specification parsing function for the i-th device. i T represents the initial data collected by the i-th device. i This represents the timestamp weight coefficient corresponding to the i-th device.

[0072] When initial data is acquired from multi-source monitoring devices, it involves different types of sensors, such as temperature sensors, humidity sensors, and pressure sensors, which generate data in different formats. Temperature sensors output data in JSON format, containing the fields "temp" and "time", such as {"temp": 25.5", "time": "2025-10-15 08:00:00"}, while humidity sensors output data in CSV format, such as "humidity, 65, 2025-10-15 08:00:00". To address these format differences, the pre-defined interface specification aims to parse different formats into a unified data structure. During implementation, a parser can be designed to identify the input format based on the device type, extract the timestamp and value, and generate a standardized initial data set, such as [{timestamp: "2025-10-15 08:00:00", value: 25.5, type: "temperature"}, {timestamp: "2025-10-15 08:00:00", value: 65, type: "humidity"}]. This facilitates subsequent processing. This standardized parsing ensures data consistency and improves the system's compatibility with multi-source data.

[0073] Step S120: To address the format differences in the initial dataset, a conversion module is used to process the standardized parsed initial data into a unified format, generating a unified data sequence containing time information and numerical markers.

[0074] A unified data sequence containing time information and numerical markers is derived using the following formula:

[0075] D u (t)=α·T norm (D raw ,t)+β·V tag (D raw (2)

[0076] In formula (2), D u (t) represents the data sequence after uniform format processing, T norm (D raw ,t) represents the initial data D raw The time information after standardized parsing, V tag (D raw ) indicates starting from the initial data D raw The extracted numerical markers, α and β represent the weighting coefficients of time information and numerical markers, respectively.

[0077] To address the format differences in the initial dataset, the conversion module needs to unify the data format. For example, timestamps vary depending on the device; some are "YYYY-MM-DDHH:MM:SS," while others are Unix timestamps such as 1728960000. The conversion module unifies all timestamps to a standard format, such as the ISO8601 format "2025-10-15T08:00:00Z." The conversion module checks the timestamp field; if it is a Unix timestamp, it converts it to the standard format using a time conversion algorithm; if the numerical units are inconsistent, such as temperatures in Celsius and Fahrenheit, the module converts Fahrenheit to Celsius, for example, 68℉ to 20℃. This unified processing generates a unified data sequence containing time information and numerical markers, such as [{time: "2025-10-15T08:00:00Z", value: 20, type: "temperature"}], ensuring consistent data format for easy analysis and storage.

[0078] The generation of standardized data sequences can significantly improve data processing efficiency. In environmental monitoring scenarios, assuming the analysis of temperature and humidity data in a factory workshop is required, without standardization and standardized processing, subsequent analysis may lead to errors due to inconsistent formats. In one embodiment, standardized data sequences can be directly input into a data analysis system to generate real-time monitoring reports, such as displaying alarms when the temperature exceeds 30°C or the humidity is below 40%. This standardized format also facilitates data fusion, for example, combining temperature and humidity data to analyze their impact on equipment operation. Preferably, standardized data sequences can be stored in a database, supporting historical data queries and trend analysis, significantly improving data utilization and system reliability.

[0079] Furthermore, the ultra-deep underground space remote monitoring method based on multi-source data fusion proposed in this embodiment includes step S200 as follows:

[0080] Step S210: Based on the unified data sequence, the K-means clustering algorithm is used to group the devices by similarity, generating a subset of groups containing similar device types.

[0081] The following formula is used to minimize the sum of squared errors of device data within a cluster, thus achieving device similarity grouping:

[0082]

[0083] In formula (3), J represents the objective function of K-means clustering, k represents the number of clusters, and C i Let x represent the i-th cluster, and let μ represent the device data points within that cluster. i This represents the center point of the i-th cluster.

[0084] When grouping devices based on similarity using a unified data sequence, the K-means clustering algorithm is employed for device classification. The unified data sequence contains time information and numerical labels, such as temperature and humidity data. The principle of K-means clustering is to iteratively calculate and divide devices into several groups based on data characteristics, with devices within each group exhibiting similar numerical patterns. For example, assuming 50 sensors collect unified data sequences showing that some sensors have temperature values ​​concentrated between 20-25℃ and humidity between 60-70%, while others have temperatures between 28-32℃ and humidity between 40-50%. K-means clustering divides the devices into two groups: one group for "low temperature and high humidity" devices, and the other group for "high temperature and low humidity" devices. In implementation, the number of clusters K should be preset to 2. The K-means clustering algorithm iteratively optimizes based on the Euclidean distance of the numerical labels, generating subsets of the groups, such as subset 1 containing 20 devices and subset 2 containing 30 devices. This grouping facilitates the identification of similarities in the operating environments of the devices.

[0085] Step S220: Based on the group subsets, obtain the location information of each group of devices, and combine the location information with the group subsets through the location embedding module to generate an intermediate dataset with location annotations.

[0086] The intermediate dataset with location annotations is obtained using the following formula:

[0087] M i ={(f i,k E pos (d i,k ))|d i，k ∈S i k1 = 1, 2, ..., n i} (4)

[0088] In formula (4), M i f represents the intermediate dataset with location labels generated in the i-th group. i，k E represents the feature vector of the k-th device in the i-th group. pos (d i，k S represents the location embedding vector of the k-th device in the i-th group. i Let n represent the i-th subset group. i This indicates the total number of devices in the group.

[0089] The location embedding vector of the k-th device in the i-th group is obtained by the following formula:

[0090] E pos (d i，k ) = W e ·concat(sin(ω·p i,k ),cos(ω·p i,k (5)

[0091] In formula (5), E pos (d i，k W represents the location embedding vector of the k-th device in the i-th group. e p represents the weight matrix for position embedding. i，k Let ω represent the position coordinate vector of the k-th device in the i-th group, ω represent the frequency parameter of the position encoding, and concat represent the vector concatenation operation.

[0092] After obtaining the location information of each group of devices, the location is combined with the group subsets through a location embedding module. Location information includes the device's geographic coordinates or relative position within the workshop, such as "Area A, coordinates (10, 20)". The principle of the location embedding module is to convert the location information into a processable vector or label and embed it into the group subsets. For example, the devices in subset 1 might be located in Area A near the ventilation openings in the workshop, and the devices in subset 2 might be located in Area B near the heating equipment. The location embedding module appends the coordinate information to the subset data, generating an intermediate dataset, such as [{time:“2025-10-15T08:00:00Z”, value:22, type:“temperature”, location:“Area A, (10, 20)”}]. This location labeling provides spatial dimension support for subsequent data integration.

[0093] Step S230: If the location labels in the intermediate dataset are complete, the time information, numerical labels and location labels are merged through the data integration module to obtain an integrated dataset with labels.

[0094] The labeled ensemble dataset is derived using the following formula:

[0095]

[0096] In formula (6), D integrated T represents the final labeled ensemble dataset. i V represents the time information of the i-th data item. i L represents the numerical label of the i-th data item. i This represents the position label of the i-th data item, and n represents the total number of data items in the intermediate dataset.

[0097] If the intermediate dataset has complete location annotations, the data integration module will merge the time information, numerical labels, and location annotations to generate an integrated dataset. The principle of the data integration module is to merge multi-dimensional data into a unified structure through association rules. For example, for the devices in subset 1, the data integration module merges the timestamp "2025-10-15T08:00:00Z", temperature value 22, and location "Region A, (10, 20)" into a single record, such as [{time: "2025-10-15T08:00:00Z", value: 22, type: "temperature", location: "Region A, (10, 20)"}]. Similarly, the data in subset 2 is also integrated into a record containing complete annotations. The integrated dataset facilitates the analysis of device operating patterns at specific locations and times, such as identifying whether devices near ventilation openings need adjustment due to low temperature and high humidity. This fusion improves data usability.

[0098] Furthermore, in the ultra-deep underground space remote monitoring method based on multi-source data fusion proposed in this embodiment, step S300 includes:

[0099] Step S310: Based on the integrated dataset, obtain environmental variable change data from the environmental variables, calculate the Pearson correlation coefficient between the environmental variable change and the equipment status through the correlation analysis module, and obtain the correlation coefficient matrix. The environmental variable change data includes temperature change data, humidity change data, and light intensity change data.

[0100] The Pearson correlation coefficient between changes in environmental variables and equipment status is calculated using the following formula:

[0101]

[0102] In formula (7), r xy x represents the correlation coefficient. i Let y represent the data point showing the change in the i-th environmental variable. i This represents the i-th device status data point. This represents the mean of the environmental variable change data. n represents the mean of the device status data, and n1 represents the total number of data samples.

[0103] The correlation coefficient matrix can be obtained using the following formula:

[0104]

[0105] Where R represents the correlation coefficient matrix, r ij This represents the correlation coefficient between the i-th environmental variable and the j-th device status parameter. The rows of the matrix represent different environmental variables, including temperature, humidity, and light intensity, and the columns represent different device status parameters.

[0106] Based on an integrated dataset that combines location and numerical labels of equipment group subsets, such as data on low-temperature, high-humidity equipment in area A of the workshop. Environmental variable variation data are obtained, including temperature fluctuations from 21°C to 24°C, humidity decreases from 65% to 62%, and light intensity increases from 500 lux to 600 lux. This variation data is collected in real-time by sensors and correlated with equipment status such as operating power and failure rate.

[0107] The correlation analysis module calculates the Pearson correlation coefficient between changes in environmental variables and equipment status, generating a correlation coefficient matrix. For example, the correlation coefficient between temperature change and equipment power is 0.75, the correlation coefficient between humidity change and failure rate is -0.68, and the correlation coefficient between light intensity change and power is 0.45. The principle of the Pearson correlation coefficient is to measure the strength of the linear relationship through covariance standardization; a positive value indicates a positive correlation, and a negative value indicates an inverse correlation. This calculation facilitates the quantification of the impact of environmental factors on equipment.

[0108] Step S320: If at least one coefficient in the correlation coefficient matrix exceeds a preset threshold, the environmental variable change data and equipment status data are weighted and combined through the matrix fusion module to generate a dynamic matrix.

[0109] The following formula is used to generate a dynamic matrix by weighted combination of environmental variables and device status data:

[0110]

[0111] In formula (8), D ij (t) represents the value of the element in the i-th row and j-th column of the dynamic matrix at time t, w kl E represents the fusion weight between environmental variable k and device state l. k (t) represents the k-th environmental variable data at time t, S l (t) represents the l-th device state data at time t, q represents the number of environmental variables involved in the fusion, r represents the number of device states involved in the fusion, and β ij Indicates the bias term

[0112] If at least one coefficient in the correlation coefficient matrix exceeds a preset threshold, such as 0.7, the matrix fusion module is triggered. For example, if the temperature change coefficient exceeds the threshold of 0.75, the module performs a weighted combination of environmental variable change data and equipment status data, with the weights based on the absolute values ​​of the coefficients, such as a temperature weight of 0.75 and a humidity weight of 0.68, generating a dynamic matrix. This dynamic matrix dynamically reflects real-time correlations, such as the power increase pattern of equipment in region A when the temperature rises.

[0113] Step S330: Use principal component analysis algorithm to reduce the dimensionality of the dynamic matrix, extract the main distribution features, and obtain the dimensionality-reduced feature matrix.

[0114] The linear transformation from a high-dimensional space to a low-dimensional principal component space can be achieved using the following formula:

[0115] Y = W T (X-μ) (9)

[0116] In formula (9), Y represents the feature matrix after dimensionality reduction, W represents the transformation matrix composed of the first m principal component vectors, X represents the original dynamic matrix, μ represents the mean vector of the original data, and T represents the transpose.

[0117] Principal Component Analysis (PCA) algorithms reduce the dimensionality of dynamic matrices and extract key distribution features. PCA utilizes the principle of variance maximization to merge highly correlated dimensions and reduce noise. For example, an original 10-dimensional matrix can be reduced to 3 dimensions, with key features including integrated environmental fluctuations and equipment response patterns, resulting in a dimensionality-reduced feature matrix that facilitates efficient processing of large-scale data.

[0118] Step S340: The data points of the feature matrix are mapped to the preset monitoring grid through the distribution mapping module to generate a monitoring matrix that reflects the distribution relationship between environmental variables and equipment status.

[0119] The monitoring matrix reflecting the relationship between environmental variables and equipment status distribution is derived using the following formula:

[0120]

[0121] In formula (10), M ij F represents the value of the element in the i-th row and j-th column of the monitoring matrix. kl Let d represent the l-th eigenvalue of the k-th data point in the feature matrix, G represent the Gaussian kernel function used to calculate the mapping weights from the data points to the grid, and d represent the eigenvalue of the l-th data point in the feature matrix. kij σ represents the distance from the k-th data point to the grid position (i, j). ij The bandwidth parameter represents the grid position (i,j), and K represents the total number of data points in the feature matrix.

[0122] The distribution mapping module maps feature matrix data points to preset monitoring grids, such as 10x10 grids corresponding to location coordinates. For example, feature points of low-temperature, high-humidity equipment in region A are mapped to the vicinity of grid (5, 5), generating a monitoring matrix. This monitoring matrix uses color coding to represent the intensity of distribution relationships, with high-correlation areas highlighted in red, facilitating visualization of the spatiotemporal distribution of environmental and equipment status. In region B, with high-temperature, low-humidity equipment, if the light intensity variation coefficient exceeds a threshold, the dimensionality reduction features after dynamic matrix fusion highlight the overheating risk, mapping to grid (8, 8). The monitoring matrix reveals concentrated anomalies in the heating zone, supporting targeted intervention. This multi-module linkage improves the accuracy of anomaly detection.

[0123] Furthermore, in the ultra-deep underground space remote monitoring method based on multi-source data fusion proposed in this embodiment, step S400 includes:

[0124] Step S410: Based on the monitoring matrix, extract the environmental variable trend sequence through a time series filter, calculate the difference value of consecutive time points in the environmental variable trend sequence, and if the difference value exceeds a preset threshold, mark the potential change point to obtain the trend sequence of the marked change point.

[0125] The following formula describes the trend sequence of environmental variables extracted from the original monitoring matrix using a weighted average method:

[0126]

[0127] In formula (11), Y t h represents the trend value of the environmental variable at time t after processing by the time series filter. k X represents the weighting coefficients of the filter at a lag of k time steps. t-k K1 represents the raw monitoring data at time tk, and K1 represents the order of the filter.

[0128] The absolute difference of a trend sequence between consecutive time points is calculated using the following formula, which is used to quantify the magnitude of change of environmental variables at adjacent time points:

[0129] D t =|Y t -Y t-1 | (12)

[0130] In formula (12), D t Y represents the difference at time t. t Y represents the trend value of the environmental variable at time t after processing by a time series filter. t-1 This represents the trend value of the environmental variable at time t-1 after processing by the time series filter.

[0131] The following formula defines the criteria for detecting change points: when the difference value exceeds a preset threshold, it is marked as a potential change point:

[0132]

[0133] In formula (13), C t This indicates the status of the change point marker at time t. A value of 1 indicates that a potential change point has been detected, and a value of 0 indicates that no change point has been detected. (D) t θ represents the difference value at time t, and θ represents the preset threshold parameter.

[0134] To extract the trend sequence of environmental variables from the monitoring matrix, a time series filter is used for smoothing to capture long-term trends. Time series filters are typically based on the sliding window mean method, aiming to reduce short-term noise interference and highlight the continuous variation patterns of environmental variables such as temperature and humidity. Temperature data from region A is processed by the filter to generate a trend sequence showing a slow rise in temperature from 22℃ to 25℃. Next, the differences between consecutive time points are calculated; for example, the temperature difference between adjacent time points is 0.5℃ or 0.7℃. If a preset threshold of 0.6℃ is set, the difference of 0.7℃ is marked as a potential change point, resulting in a trend sequence containing these change points. This method facilitates the identification of abrupt changes in environmental variables, providing a foundation for subsequent analysis.

[0135] Step S420: Use a support vector machine to classify the trend sequence of labeled change points, distinguish between potential change points and normal points, and generate a classification label sequence.

[0136] Define a decision formula for generating the sequence of class labels, which converts the continuous output of the support vector machine into discrete class labels:

[0137]

[0138] In formula (14), y k G represents the category label at the k-th time point. k This represents the classification score of the support vector machine for the k-th data point, where sign represents the sign function; when g k When g is greater than or equal to 0, the label is +1, indicating that the point is a potential point of change. k When the value is less than 0, the label is -1, indicating that the point is a normal point. Finally, a complete classification label sequence is formed to identify the change pattern in the trend sequence.

[0139] Support Vector Machines (SVMs) are used to classify trend sequences, distinguishing between potential points of change and normal points. The SVM constructs a hyperplane to separate change points from normal points, generating a sequence of classification labels. For example, temperature changes in region A are labeled as "abnormal," while humidity changes are labeled as "normal" because their difference values ​​do not exceed a threshold. The classification label sequence visually reflects the abnormal patterns of environmental variables, facilitating further analysis of abnormal distributions.

[0140] Step S430: Obtain the dense region of abnormal labels from the classification label sequence, determine the boundary of the high-risk region through clustering algorithm, and if the abnormal density in the high-risk region exceeds the preset threshold, fuse the boundary information to obtain the outline of the high-risk region.

[0141] The following formula is used to calculate the density of outlier labels at each position in the classification label sequence:

[0142]

[0143] In formula (15), D i N represents the density of anomalous labels at position i. r (i) represents the neighborhood with radius r centered at position i, I(L j =1) indicates that the value is 1 when the label at position j is an anomaly, and 0 otherwise; |N r (i)| represents the total number of positions in the neighborhood.

[0144] The cluster boundaries of high-risk regions are determined by minimizing the weighted sum of squared distances using the following formula:

[0145]

[0146] In formula (16), C k Denotes the set of k cluster centers. This indicates finding the cluster center partition C, x that minimizes the following expression. i μ represents the feature vector of the i-th abnormal region. j Let w represent the j-th cluster center. ij n represents the weight of sample i belonging to cluster j, n3 represents the total number of outlier regions, and k3 represents the total number of cluster centers.

[0147] The complete outline is obtained by merging the boundaries of all regions that meet the density condition using the following formula:

[0148]

[0149] In formula (17), B final ρ represents the final fused high-risk region boundary outline. kθ represents the abnormal density of the k-th region, θ1 represents the preset abnormal density threshold, and φ k (x,y) represents the boundary function of the k-th region, when φ k (x, y) = 0 represents a boundary point.

[0150] The algorithm extracts densely populated regions of anomalous labels from the classification label sequence and uses the DBSCAN clustering algorithm to determine the boundaries of high-risk regions. DBSCAN is based on density clustering and is suitable for identifying irregularly shaped densely populated anomalous regions. For example, if the temperature anomaly labels in region A are concentrated around grid (5, 5), the clustering algorithm calculates the anomaly point density and identifies a high-risk region centered at (5, 5) with a radius of 2 grid units. If the anomaly density within the region exceeds a preset threshold, such as 5 anomaly points per square grid, the boundary information is fused to generate a high-risk region outline. This outline clearly defines the concentrated anomaly region, facilitating accurate problem localization.

[0151] Step S440: Overlay the outline of the high-risk area onto the corresponding position of the monitoring matrix using a raster rendering tool, mark the position of the abnormal mode, and generate a mapping layer with annotated anomalies.

[0152] The mapping layer with anomalies is obtained using the following formula:

[0153]

[0154] In formula (18), A label (i,j) represents the anomaly labeling status at position (i,j) in the monitoring matrix. A value of 1 indicates that an anomaly pattern exists at that position and needs to be labeled, while a value of 0 indicates a normal state; D anomaly (i, j) represents D anomaly The anomaly detection value of (i,j), where θ2 represents the threshold parameter for anomaly determination.

[0155] A raster rendering tool overlays the outlines of high-risk areas onto the monitoring matrix, generating a mapping layer that annotates anomalies. The raster rendering tool uses color coding for anomalous areas; for example, high-risk areas are marked in red, and normal areas in green. The temperature anomaly outline in area A is displayed in red near the grid (5, 5), indicating that the device may face overheating risk due to temperature fluctuations. This visualization layer intuitively shows the relationship between anomaly locations and environmental variables, supporting rapid decision-making and intervention.

[0156] Furthermore, the ultra-deep underground space remote monitoring method based on multi-source data fusion proposed in this embodiment includes step S500 as follows:

[0157] Step S510: Based on the mapping layer, extract the coordinate data of the high-risk area outline, and use a data encoding tool to serialize the coordinate data to obtain a protocol data packet containing area information.

[0158] The mathematical process of serializing coordinate data using data encoding tools is described by the following formula:

[0159]

[0160] In formula (19), S(C) represents the encoding result after serialization, and E enc This represents the data encoding function, where C represents the input coordinate data, and K represents the input coordinate data. serial c represents the serialization encoding key. k B represents the coordinate data value of the k-th position. k L represents the base of the k-th bit encoding, and L represents the total length of the coordinate data.

[0161] The following formula represents the process of assembling serialized region coordinate information into a standard protocol data packet:

[0162]

[0163] In formula (20), P protocol H represents the final generated protocol data packet. header Indicates data packet header information, D region T represents the data payload portion containing regional information. tail This indicates the checksum information at the end of the data packet. This indicates the packet assembly operator.

[0164] To extract coordinate data from the contours of high-risk areas in the mapping layer, a raster scanning algorithm was used to traverse the red-marked areas and record the coordinate sequence of boundary points. Specifically, the temperature anomaly contour of region A was centered at grid (5, 5), and coordinates such as (4, 5) and (5, 6) were extracted to form a closed polygonal path. This coordinate data captures the geometry of the anomaly concentration, facilitating subsequent transmission and analysis.

[0165] Coordinate data is serialized using data encoding tools, converting floating-point coordinates into a compact binary format to generate protocol data packets. For example, the coordinate sequence is compressed into a byte stream after JSON encoding, with the packet header containing a region ID such as "A-High Temperature," and the packet body embedding a coordinate array and a density value of 5 points / grid. This serialization ensures the efficiency and integrity of data during network transmission, supporting real-time sharing of high-risk information.

[0166] Step S520: The information flow in the protocol data packet is fragmented by the computing node. If the number of fragments exceeds the preset threshold, the fragments are reassembled using a parallel processing tool to obtain the reassembled information flow sequence.

[0167] The following formula is used to calculate the total number of fragments generated when protocol data packets are fragmented by compute nodes:

[0168]

[0169] In formula (21), F s D represents the number of fragments. p S represents the total size of the protocol data packets. u C represents the standard size of a single fragment. n This represents the processing coefficient of the computing node.

[0170] The following formula is used to determine whether the current number of shards exceeds a preset threshold. When the number of shards exceeds the threshold, a parallel processing mechanism is triggered:

[0171] T h =max(F s θ preset ) (twenty two)

[0172] In formula (22), T h F represents the threshold judgment result. s Indicates the current number of fragments, θ preset This indicates the preset threshold for the number of fragments.

[0173] The following formula describes the process by which a parallel processing tool reassembles multiple fragments to obtain a complete information stream sequence:

[0174]

[0175] In formula (23), R seq P represents the recombined information flow sequence. tool The recombination algorithm for parallel processing tools, F1 to F... k Representing each fragment, This indicates the parallel recombination operator.

[0176] When computing nodes fragment the protocol data packet flow, they first assess the packet size. If it exceeds a threshold such as 1MB, it is evenly divided into multiple fragments. The protocol packet size in region A is 1.2MB, and it is divided into three 512KB fragments, each carrying partial coordinates and a checksum. If the number of fragments exceeds the threshold of 2, a parallel processing tool is triggered to reassemble them. By merging fragments through multi-threading, the original information flow sequence is restored, avoiding data loss due to transmission interruptions.

[0177] Parallel processing tools utilize distributed queue scheduling for fragment reassembly, prioritizing high-priority abnormal data. Nodes synchronously verify fragment hash values ​​to ensure the reassembled sequence matches the original packet; for example, the reconstructed sequence for region A contains 20 coordinate points. This approach improves the robustness of processing large-scale monitoring data.

[0178] Step S530: Use a convolutional neural network to extract features from the recombined information flow sequence, obtain optimized signal features, and obtain a signal sequence containing the correspondence between variables.

[0179] The optimized signal characteristics are obtained using the following formula:

[0180]

[0181] In formula (24), F opt (x) represents the optimized signal feature output, K3 represents the number of convolutional kernels, and W... k Let x represent the weights of the k-th convolutional kernel, σ represent the activation function, M3 and N3 represent the height and width dimensions of the input information stream sequence, respectively, and x i,j h represents the value of the element at position (i,j) in the input sequence. k,i,j b represents the parameters of the k-th convolutional kernel at position (i, j). k This represents the bias term of the k-th convolution kernel.

[0182] A convolutional neural network is used to extract features from the reconstructed information stream sequence. Multiple convolutional layers scan the sequence to capture spatial patterns and variable relationships between coordinates. The input sequence is processed by a 1D convolutional filter to extract edge features of temperature anomalies, outputting optimized signal features such as a peak intensity of 0.8. The final signal sequence contains a mapping between temperature variables and coordinates, such as (5, 5) corresponding to a 25℃ peak. This feature optimization enhances the quantitative representation of anomaly patterns, supporting accurate risk assessment and intervention decisions.

[0183] Furthermore, in the ultra-deep underground space remote monitoring method based on multi-source data fusion proposed in this embodiment, step S600 includes:

[0184] Step S610: For the signal sequence, use a time series prediction model to estimate future trend information, extract the peak point coordinates from the future trend information, and obtain a preliminary trend curve.

[0185] The following formula is used to extract peak point coordinates from the predicted trend:

[0186]

[0187] In formula (25), P k Let t represent the set of the k-th peak points. kThe time coordinate of the peak point is represented by f(t), the predicted trend function is represented by f(t), the second derivative being less than zero ensures that the point is a maximum point, δ represents the neighborhood range, and the condition ensures that the point is a maximum value in the local neighborhood.

[0188] The method for constructing the initial trend curve is expressed by the following formula:

[0189]

[0190] In formula (26), C(t) represents the final trend curve, w k B represents the weighting coefficients of the basis functions. k (t) represents the k-th basis function, m5 represents the total number of basis functions, α5 represents the peak point influence factor, and β i P represents the weight of the i-th peak point. i (t) represents the influence function of the i-th peak point, and n5 represents the total number of peak points.

[0191] A time-series prediction model is used to estimate future trend information for the signal sequence, and the peak point coordinates are extracted to obtain a preliminary trend curve. The temperature signal sequence of region A is input into an LSTM model, and the trend for the next hour is predicted using historical 25℃ peak data. The peak coordinates, such as (6, 7), correspond to the predicted 28℃ point, forming a preliminary curve path. This prediction captures temperature rise patterns, facilitating early intervention for high-temperature risks. The time-series prediction model scans the signal sequence one frame every 5 minutes, and the output curve contains 5 peak points, supporting real-time trend visualization.

[0192] Step S620: Based on the preliminary trend curve, a clustering algorithm is used to group the peak point locations. If the group density exceeds a preset threshold, noise points are removed using a density filtering tool to determine the key point location sequence.

[0193] The following formula is used to calculate the density distribution around each peak point:

[0194]

[0195] In formula (27), D k P represents the local density value at the k-th peak point. k This represents the coordinates of the k-th peak point. Indicated by P k Let P be the set of peak points within a neighborhood with a center radius of r. j Let represent the j-th peak point in the neighborhood, and σ represent the bandwidth parameter of the Gaussian kernel function.

[0196] The following formula is used to assign peak points to the group containing the nearest cluster center:

[0197]

[0198] In formula (28), C i Indicates the cluster label to which the i-th peak point belongs, x i μ represents the eigenvector of the i-th peak point. j Let represent the location vector of the j-th cluster center, d represent the Euclidean distance function, and m7 represent the dimension of the feature vector.

[0199] The following formula is used to filter out the true critical peak locations based on a density threshold:

[0200]

[0201] In formula (29), F(p) represents the filtering result of point p, ρ(p) represents the density value at point p, and θ3 represents the preset density threshold. When F(p) equals 1, it means that the point is retained as a key point. When F(p) equals 0, it means that the point is identified as a noise point and is filtered out.

[0202] The peak points are grouped using a clustering algorithm based on the initial trend curve. If the group density exceeds a preset threshold, noise points are removed using a density filtering tool to determine the key point sequence. For example, the K-means algorithm divides five points (6, 7), (5, 8), etc., into two groups. The density of group 1 reaches 0.9, exceeding the threshold of 0.8, so one noise point is filtered out, retaining four key points such as (6, 7) and (5, 8). This grouping focuses on abnormally concentrated areas, improving sequence accuracy.

[0203] Step S630: Based on the key point location sequence, use a rasterization mapping tool to convert the peak point location into a two-dimensional coordinate grid, obtain the risk intensity value of each grid from the two-dimensional coordinate grid, and obtain the risk distribution matrix.

[0204] The risk intensity value for each grid is derived using the following formula:

[0205]

[0206] In formula (30), R mn This represents the risk intensity value at grid location m rows and n columns, α7 represents the risk propagation coefficient, and x m and y n These represent the x and y coordinates of the target grid, respectively. p and y p The x and y coordinates of the peak point represent the position of the peak point, σ1 represents the standard deviation parameter of the risk impact range, and I represent the standard deviation parameter of the risk impact range. p This represents the original intensity value at the peak point.

[0207] The risk distribution matrix is ​​derived using the following formula:

[0208]

[0209] In formula (31), M h This represents the final generated risk distribution matrix, where the number of rows M and columns N correspond to the number of rows and columns of the two-dimensional coordinate grid, respectively, and R... ij This represents the risk intensity value corresponding to the element in the i-th row and j-th column of the matrix. The entire matrix fully describes the spatial distribution characteristics of risk intensity within the study area.

[0210] Based on the sequence of key point locations, a rasterization mapping tool is used to convert them into a two-dimensional coordinate grid. From the grid, the risk intensity value of each grid cell is obtained to obtain the risk distribution matrix. For example, based on the grid (5, 5), 4 key points are mapped to 9 grid cells, with intensity values ​​such as 2.5 for the center grid and 1.2 for the edges, forming a 5x5 matrix. This rasterization quantifies the spatial distribution and supports matrix analysis.

[0211] Step S640: For the risk distribution matrix, the risk intensity value is converted into a monitoring status label using a status mapping table. If the monitoring status label meets the preset level, the status level of each area is determined to obtain the risk distribution map of the corresponding monitoring status.

[0212] The defined state mapping table transformation formula is:

[0213]

[0214] In formula (32), S i R represents the monitoring status label of the i-th region. i Represents the risk intensity value of the i-th region, T1 to T k-1 Representing different threshold boundaries, L1 to L k The corresponding status level label is represented by a piecewise function, which discretizes the continuous risk intensity values ​​into specific monitoring status labels.

[0215] The predefined formula for determining the default level is:

[0216]

[0217] In formula (33), P ij S represents the judgment result at position (i, j). ij S represents the monitoring status label value at location (i, j). preset This indicates a preset level threshold. When the monitoring status label reaches or exceeds the preset level, output 1 to indicate that the condition is met; otherwise, output 0 to indicate that the condition is not met.

[0218] The defined formula for generating the risk distribution map is:

[0219]

[0220] In formula (34), M ij This represents the displayed value at position (i, j) in the risk distribution map, K5 represents the total number of state levels, and δ k I(S) represents the display weight or color code value corresponding to the k-th state level. ij =L k ) indicates that the indicator function is in operation when the state label at position (i, j) equals L. k The result is 1 if the value is 1 otherwise, and the final risk distribution map is obtained by weighted summation to show the effect.

[0221] A status mapping table is used to convert risk intensity values ​​into monitoring status labels for the risk distribution matrix. If a label matches a preset level, the status level of each area is determined, resulting in a corresponding monitoring status risk distribution map. For example, the status mapping table defines an intensity >2 as a "high temperature alarm" label. If 8 grids in the risk distribution matrix match level 3, area A is judged as high risk, and a distribution map with 9 red cells is generated. This conversion intuitively displays the status, facilitating rapid operator response to anomalies.

[0222] Please see Figure 2 This invention provides a remote monitoring system for ultra-deep underground space based on multi-source data fusion, used to execute the aforementioned remote monitoring method for ultra-deep underground space based on multi-source data fusion. It includes a unified data sequence acquisition module 10, an integrated dataset acquisition module 20, a monitoring matrix acquisition module 30, a mapping layer acquisition module 40, a signal sequence acquisition module 50, and a risk distribution map acquisition module 60. The unified data sequence acquisition module 10 acquires initial data from multi-source monitoring devices through a preset interface specification, processes different format differences using a conversion module, and obtains a unified data sequence containing time information and numerical labels. The integrated dataset acquisition module 20, based on the unified data sequence, uses a clustering algorithm to group similar device types, determines group subsets, and embeds location information within the group subsets to obtain a map with labels. The integrated dataset includes: a monitoring matrix acquisition module 30, which acquires correlation information of environmental variable changes for the integrated dataset; if the coefficients between correlation information exceed a threshold, they are fused into a dynamic matrix to obtain a monitoring matrix reflecting the distribution; a mapping layer acquisition module 40, which extracts trend information from the monitoring matrix, uses a classification model to identify abnormal patterns, determines high-risk areas, and obtains a mapping layer annotating anomalies; a signal sequence acquisition module 50, which generates transmission protocol data based on the mapping layer, uses computing nodes to process the information flow within the transmission protocol data, and obtains an optimized signal sequence, which integrates the correspondence of variables; and a risk distribution map acquisition module 60, which uses a prediction model to estimate trend information for the signal sequence, determines the location of key points in the trend information, and obtains a risk distribution map, which corresponds to the monitoring status.

[0223] This embodiment provides a remote monitoring method and system for ultra-deep underground space based on multi-source data fusion. Compared with existing technologies, it acquires initial data from multiple sources through a preset interface, which is then converted into a sequence containing time and numerical values ​​by a conversion module. A clustering algorithm then groups similar devices and embeds their locations, generating a labeled integrated dataset. Environmental variables are then correlated, and if the coefficients exceed a threshold, they are fused into a dynamic monitoring matrix. Trends are extracted from this matrix, and anomalies are identified using a classification model, outputting a labeled mapping layer. Transmission protocol data is generated based on this matrix and optimized by computational nodes into a signal sequence integrating variable relationships. Finally, a prediction model estimates key trend points, generating a risk distribution map corresponding to the monitoring status. The most important invention lies in the fusion mechanism of the dynamic monitoring matrix and signal sequence, achieving a closed-loop flow of data from heterogeneous to optimized data, thereby accurately predicting high-risk distributions and improving environmental monitoring response efficiency. The overall technical effect is to construct a real-time, accurate risk early warning system, significantly reducing losses from emergencies.

[0224] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A remote monitoring method for ultra-deep underground space based on multi-source data fusion, characterized in that, Includes the following steps: S100. Initial data is obtained from multi-source monitoring devices through a preset interface specification. Different format differences are processed by a conversion module to obtain a unified data sequence. The unified data sequence includes time information and numerical markers. S200. Based on the unified data sequence, a clustering algorithm is used to group similar device types, determine the group subsets, and embed location information in the group subsets to obtain an integrated dataset with annotations. S300. For the integrated dataset, obtain the correlation information of environmental variable changes. If the coefficients between the correlation information exceed the threshold, they are fused into a dynamic matrix to obtain a monitoring matrix that reflects the distribution. S400. Extract trend information from the monitoring matrix, identify abnormal patterns using a classification model, determine high-risk areas, and obtain a mapping layer that marks anomalies. S500. Based on the mapping layer, generate transmission protocol data, use computing nodes to process the information flow within the transmission protocol data, and obtain an optimized signal sequence. The signal sequence integrates the correspondence between variables. S600. For the signal sequence, a prediction model is used to estimate trend information, determine the location of key points in the trend information, and obtain a risk distribution map, which corresponds to the monitoring status.

2. The method for remote monitoring of ultra-deep underground space based on multi-source data fusion according to claim 1, characterized in that, Step S100 includes: S110. Obtain initial data from multi-source monitoring equipment, and standardize and parse the initial data using a preset interface specification to obtain an initial data set containing timestamps and values. The initial dataset, containing timestamps and numerical values, is derived using the following formula: Among them, D s This represents the initial data set containing timestamps and values, where M represents the total number of monitoring devices, and P... i R represents the default interface specification parsing function for the i-th device. i T represents the initial data collected by the i-th device. i This represents the timestamp weight coefficient corresponding to the i-th device; S120. In response to the format differences in the initial data set, a conversion module is used to process the standardized parsed initial data into a unified format, generating a unified data sequence containing time information and numerical markers. A unified data sequence containing time information and numerical markers is derived using the following formula: D u (t)=α·T norm (D raw ,t)+β·V tag (D raw ) Among them, D u (t) represents the data sequence after uniform format processing, T norm (D raw ,t) represents the initial data D raw The time information after standardized parsing, V tag (D raw ) indicates starting from the initial data D raw The extracted numerical markers, α and β represent the weighting coefficients of time information and numerical markers, respectively.

3. The method for remote monitoring of ultra-deep underground space based on multi-source data fusion according to claim 1, characterized in that, Step S200 includes: S210. Based on the unified data sequence, the devices are grouped by similarity using the K-means clustering algorithm to generate a subset of groups containing similar device types; The following formula is used to minimize the sum of squared errors of device data within a cluster, thus achieving device similarity grouping: Where J represents the objective function of K-means clustering, k represents the number of clusters, and C i Let x represent the i-th cluster, and let μ represent the device data points within that cluster. i This represents the center point of the i-th cluster; S220. Based on the grouped subsets, obtain the location information of each group of devices, and combine the location information with the grouped subsets through the location embedding module to generate an intermediate dataset with location annotations. The intermediate dataset with location annotations is obtained using the following formula: M i ={(f i，k E pos (d i，k ))|d i，k ∈S i ,k1=1,2,...,n i } Among them, M i f represents the intermediate dataset with location labels generated in the i-th group. i，k E represents the feature vector of the k-th device in the i-th group. pos (d i，k S represents the location embedding vector of the k-th device in the i-th group. i Let n represent the i-th subset group. i This indicates the total number of devices in the group; The location embedding vector of the k-th device in the i-th group is obtained by the following formula: AND poa (d i，k )=W e ·concat(sin(ω·p i,k ),cos(ω·p i,k )) Among them, E pos (d i，k W represents the location embedding vector of the k-th device in the i-th group. e p represents the weight matrix for position embedding. i，k Let ω represent the position coordinate vector of the k-th device in the i-th group, ω represent the frequency parameter of the position encoding, and concat represent the vector concatenation operation; S230. If the location labels in the intermediate dataset are complete, the time information, numerical labels and location labels are fused by the data integration module to obtain an integrated dataset with labels. The labeled ensemble dataset is derived using the following formula: Among them, D integrated T represents the final labeled ensemble dataset. i V represents the time information of the i-th data item. i L represents the numerical label of the i-th data item. i This represents the position label of the i-th data item, and n represents the total number of data items in the intermediate dataset.

4. The method for remote monitoring of ultra-deep underground space based on multi-source data fusion according to claim 1, characterized in that, Step S300 includes: S310. Based on the integrated dataset, obtain environmental variable change data from environmental variables, calculate the Pearson correlation coefficient between environmental variable changes and equipment status through the correlation analysis module, and obtain the correlation coefficient matrix, wherein the environmental variable change data includes temperature change data, humidity change data and light intensity change data. The Pearson correlation coefficient between changes in environmental variables and equipment status is calculated using the following formula: Where, r xy x represents the correlation coefficient. i Let y represent the data point showing the change in the i-th environmental variable. i This represents the i-th device status data point. This represents the mean of the environmental variable change data. n1 represents the mean of the device status data, and n1 represents the total number of data samples. The correlation coefficient matrix is ​​obtained using the following formula: Where R represents the correlation coefficient matrix, r ij This represents the correlation coefficient between the i-th environmental variable and the j-th device status parameter. The rows of the matrix represent different environmental variables, including temperature, humidity, and light intensity, and the columns represent different device status parameters. S320. If at least one coefficient in the correlation coefficient matrix exceeds a preset threshold, the environmental variable change data and the equipment status data are weighted and combined by the matrix fusion module to generate a dynamic matrix. The following formula is used to generate a dynamic matrix by weighted combination of environmental variables and device status data: Among them, D ij (t) represents the value of the element in the i-th row and j-th column of the dynamic matrix at time t, w kl E represents the fusion weight between environmental variable k and device state l. k (t) represents the k-th environmental variable data at time t, S l (t) represents the l-th device state data at time t, q represents the number of environmental variables involved in the fusion, r represents the number of device states involved in the fusion, and β ij Indicates the bias term; S330. The dynamic matrix is ​​reduced in dimensionality using principal component analysis algorithm to extract the main distribution features and obtain the reduced feature matrix. The linear transformation from a high-dimensional space to a low-dimensional principal component space can be achieved using the following formula: Y=W T (X-µ) Where Y represents the feature matrix after dimensionality reduction, W represents the transformation matrix composed of the first m principal component vectors, X represents the original dynamic matrix, μ represents the mean vector of the original data, and T represents the transpose; S340. The data points of the feature matrix are mapped to a preset monitoring grid through the distribution mapping module to generate a monitoring matrix that reflects the distribution relationship between environmental variables and equipment status. The monitoring matrix reflecting the relationship between environmental variables and equipment status distribution is derived using the following formula: Among them, M ij F represents the value of the element in the i-th row and j-th column of the monitoring matrix. kl Let d represent the l-th eigenvalue of the k-th data point in the feature matrix, G represent the Gaussian kernel function used to calculate the mapping weights from the data points to the grid, and d represent the eigenvalue of the l-th data point in the feature matrix. kij σ represents the distance from the k-th data point to the grid position (i,j). ij The bandwidth parameter represents the grid position (i, j), and K represents the total number of data points in the feature matrix.

5. The method for remote monitoring of ultra-deep underground space based on multi-source data fusion according to claim 1, characterized in that, Step S400 includes: S410. Based on the monitoring matrix, extract the environmental variable trend sequence through a time series filter, calculate the difference value of consecutive time points in the environmental variable trend sequence, and if the difference value exceeds a preset threshold, mark potential change points to obtain the trend sequence of marked change points. The following formula describes the trend sequence of environmental variables extracted from the original monitoring matrix using a weighted average method: Among them, Y t h represents the trend value of the environmental variable at time t after processing by the time series filter. k X represents the weighting coefficients of the filter at a lag of k time steps. t-k This represents the raw monitoring data at time tk, and K1 represents the filter order. The absolute difference of a trend sequence between consecutive time points is calculated using the following formula, which is used to quantify the magnitude of change of environmental variables at adjacent time points: D t =|Y t -Y t-1 | Among them, D t Y represents the difference at time t. t Y represents the trend value of the environmental variable at time t after processing by a time series filter. t-1 This represents the trend value of the environmental variable at time t-1 after processing with a time series filter. The following formula defines the criteria for detecting change points: when the difference value exceeds a preset threshold, it is marked as a potential change point: Among them, C t This indicates the status of the change point marker at time t. A value of 1 indicates that a potential change point has been detected, and a value of 0 indicates that no change point has been detected. (D) t θ represents the difference value at time t, and θ represents the preset threshold parameter; S420. Use a support vector machine to classify the trend sequence of labeled change points, distinguish between potential change points and normal points, and generate a classification label sequence. Define a decision formula for generating the sequence of class labels, which converts the continuous output of the support vector machine into discrete class labels: Among them, y k G represents the category label at the k-th time point. k This represents the classification score of the support vector machine for the k-th data point, where sign represents the sign function; when g k When g is greater than or equal to 0, the label is +1, indicating that the point is a potential point of change. k When the value is less than 0, the label is -1, indicating that the point is a normal point. Finally, a complete classification label sequence is formed to identify the change pattern in the trend sequence. S430. Obtain the dense region of abnormal labels from the classification label sequence, determine the boundary of the high-risk region through a clustering algorithm, and if the abnormal density in the high-risk region exceeds a preset threshold, fuse the boundary information to obtain the outline of the high-risk region. The following formula is used to calculate the density of outlier labels at each position in the classification label sequence: Among them, D i N represents the density of anomalous labels at position i. r (i) represents the neighborhood with radius r centered at position i, I(L j =1) indicates that the value is 1 when the label at position j is an anomaly, and 0 otherwise; |N r (i)| represents the total number of positions in the neighborhood; The cluster boundaries of high-risk regions are determined by minimizing the weighted sum of squared distances using the following formula: Among them, C k Denotes the set of k cluster centers. This indicates finding the cluster center partition C, x that minimizes the following expression. i μ represents the feature vector of the i-th abnormal region. j Let w represent the j-th cluster center. ij n3 represents the weight of sample i belonging to cluster j, n3 represents the total number of outlier regions, and k3 represents the total number of cluster centers. The complete outline is obtained by merging the boundaries of all regions that meet the density condition using the following formula: Among them, B final ρ represents the outline boundary of the high-risk area after final fusion. k θ represents the abnormal density of the k-th region, θ1 represents the preset abnormal density threshold, and φ k (x, y) represents the boundary function of the k-th region, when φ k (x, y) = 0 represents a boundary point; S440. Using a raster rendering tool, the outline of the high-risk area is superimposed onto the corresponding position of the monitoring matrix, the abnormal pattern position is marked, and a mapping layer with anomalies is generated. The mapping layer with anomalies is obtained using the following formula: Among them, A label (i, j) represents the anomaly labeling status at position (i, j) in the monitoring matrix. A value of 1 indicates that an anomaly pattern exists at that position and needs to be labeled, while a value of 0 indicates a normal state; D anomaly (i, j) represents D anomaly The anomaly detection value of (i, j), where θ2 represents the threshold parameter for anomaly determination.

6. The method for remote monitoring of ultra-deep underground space based on multi-source data fusion according to claim 1, characterized in that, Step S500 includes: S510. Based on the mapping layer, extract the coordinate data of the high-risk area outline, and use a data encoding tool to serialize the coordinate data to obtain a protocol data packet containing area information. The mathematical process of serializing coordinate data using data encoding tools is described by the following formula: Where S(C) represents the encoded result after serialization, and E enc This represents the data encoding function, where C represents the input coordinate data, and K represents the input coordinate data. serial c represents the serialization encoding key. k B represents the coordinate data value of the k-th position. k represents the base of the k-th encoding bit, and L represents the total length of the coordinate data; The following formula represents the process of assembling serialized region coordinate information into a standard protocol data packet: Among them, P protocol H represents the final generated protocol data packet. header Indicates data packet header information, D region T represents the data payload portion containing regional information. tail This indicates the checksum information at the end of the data packet. Indicates the packet assembly operator; S520. The information flow in the protocol data packet is fragmented by the computing node. If the number of fragments exceeds a preset threshold, the fragments are reassembled using a parallel processing tool to obtain the reassembled information flow sequence. S530. Use a convolutional neural network to extract features from the recombined information flow sequence, obtain optimized signal features, and obtain a signal sequence containing the correspondence between variables.

7. The method for remote monitoring of ultra-deep underground space based on multi-source data fusion according to claim 6, characterized in that, In step S520, the total number of fragments generated when the protocol data packets are fragmented by the computing node is calculated using the following formula: Among them, F s D represents the number of fragments. p S represents the total size of the protocol data packets. u C represents the standard size of a single fragment. n This represents the processing coefficient of the computing node; The following formula is used to determine whether the current number of shards exceeds a preset threshold. When the number of shards exceeds the threshold, a parallel processing mechanism is triggered: T h =max(F s ,θ preset ) Among them, T h F represents the threshold judgment result. s Indicates the current number of fragments, θ preset This indicates the preset threshold for the number of fragments; The following formula describes the process by which a parallel processing tool reassembles multiple fragments to obtain a complete information stream sequence: Among them, R seq P represents the recombined information flow sequence. tool The reorganization algorithm for parallel processing tools, F1 to F... k Representing each fragment, This indicates the parallel recombination operator.

8. The method for remote monitoring of ultra-deep underground space based on multi-source data fusion according to claim 7, characterized in that, In step S530, the optimized signal characteristics are obtained using the following formula: Among them, F opt (x) represents the optimized signal feature output, K3 represents the number of convolutional kernels, and W... k Let x represent the weights of the k-th convolutional kernel, σ represent the activation function, M3 and N3 represent the height and width dimensions of the input information stream sequence, respectively, and x i，j h represents the value of the element at position (i, j) in the input sequence. k，i，j b represents the parameters of the k-th convolutional kernel at position (i, j). k This represents the bias term of the k-th convolution kernel.

9. The method for remote monitoring of ultra-deep underground space based on multi-source data fusion according to claim 1, characterized in that, Step S600 includes: S610. For the signal sequence, a time series prediction model is used to estimate future trend information, and the peak point coordinates are extracted from the future trend information to obtain a preliminary trend curve. S620. Based on the preliminary trend curve, a clustering algorithm is used to group the peak point positions. If the group density exceeds a preset threshold, noise points are removed using a density filtering tool to determine the key point position sequence. S630. Based on the key point location sequence, the peak point locations are converted into a two-dimensional coordinate grid using a rasterization mapping tool. The risk intensity value of each grid is obtained from the two-dimensional coordinate grid to obtain the risk distribution matrix. S640. For the risk distribution matrix, a state mapping table is used to convert the risk intensity value into a monitoring state label. If the monitoring state label meets the preset level, the state level of each area is determined to obtain the risk distribution map of the corresponding monitoring state.

10. A remote monitoring system for ultra-deep underground space based on multi-source data fusion, used to execute the remote monitoring method for ultra-deep underground space based on multi-source data fusion as described in any one of claims 1 to 9, characterized in that, include: The unified data sequence acquisition module (10) is used to acquire initial data from multi-source monitoring devices through a preset interface specification, and use a conversion module to process the differences in different formats to obtain a unified data sequence, wherein the unified data sequence contains time information and numerical markers; The integrated dataset acquisition module (20) is used to group similar device types according to the unified data sequence using a clustering algorithm, determine the group subsets, and embed location information in the group subsets to obtain an integrated dataset with annotations. The monitoring matrix acquisition module (30) is used to acquire the correlation information of environmental variable changes for the integrated dataset. If the coefficient between the correlation information exceeds the threshold, it is fused into a dynamic matrix to obtain a monitoring matrix that reflects the distribution. The mapping layer acquisition module (40) is used to extract change trend information from the monitoring matrix, identify abnormal patterns using a classification model, determine high-risk areas, and obtain a mapping layer with anomalies marked. The signal sequence acquisition module (50) is used to generate transmission protocol data according to the mapping layer, process the information flow in the transmission protocol data using computing nodes, and obtain an optimized signal sequence, wherein the signal sequence integrates the variable correspondence relationship; The risk distribution map acquisition module (60) is used to estimate trend information using a prediction model for the signal sequence, determine the location of key points in the trend information, and obtain a risk distribution map, wherein the risk distribution map corresponds to the monitoring status.

Images