A method and system for intelligent prediction of the stability of surrounding rock in underground coal mine roadways

Abstract

This invention discloses an intelligent prediction method and system for the stability of surrounding rock in underground coal mine roadways, relating to the field of underground coal mine surrounding rock stability monitoring technology. The method includes collecting real-time data streams related to the working conditions of the surrounding rock in the roadway, and attaching a unified timestamp and collection location identifier to each data point to form a multi-source data set with temporal consistency and spatial correlation. Simultaneously, data consistency control and cross-monitoring point correlation analysis are performed. The multi-source data set is uploaded to an intelligent agent through a data interface, and the Dify workflow orchestration engine performs node-based orchestration and dynamic scheduling execution, constructing a structured feature set with a unified time step and spatial correlation. Based on the structured feature set, a surrounding rock stability risk scoring model is constructed, and historical cases and normative rules obtained through retrieval are used as the basis for reasoning input into the DeepSeek large language model for reasoning correction, outputting the surrounding rock stability level and prediction results.

Description

Technical Field

[0001] This invention relates to the field of coal mine underground surrounding rock stability monitoring technology, specifically to an intelligent prediction method and system for the stability of surrounding rock in underground coal mine roadways. Background Technology

[0002] With the advancement of intelligent coal mine construction, the technology for monitoring and predicting the stability of surrounding rock in underground roadways is gradually shifting from traditional experience-based judgment to data-driven and intelligent analysis. Existing technologies primarily rely on multi-source monitoring methods such as surrounding rock displacement, anchor bolt axial force, roof delamination, and microseismic activity to achieve real-time perception of surrounding rock deformation and stress state. These methods are then combined with statistical analysis, machine learning models, or numerical simulations to predict surrounding rock stability. Simultaneously, with the development of big data and artificial intelligence technologies, some studies have begun to explore the fusion analysis of multi-source monitoring data and the introduction of deep learning or expert systems to improve prediction accuracy. Furthermore, knowledge-based retrieval-enhanced analysis methods are increasingly being applied to engineering decision support, providing a certain technical foundation for assessing the stability of surrounding rock under complex geological conditions.

[0003] However, existing methods for predicting surrounding rock stability still have shortcomings. At the data level, multi-source heterogeneous monitoring data (such as displacement, stress, microseismic, and seepage data) typically originate from different sensing systems, lacking a unified time calibration and spatial correlation mechanism. This makes it difficult to effectively fuse and collaboratively analyze the data, and to accurately reflect the multi-physics coupling characteristics of the surrounding rock. Regarding data processing and analysis workflows, existing methods often rely on fixed procedures or single models, lacking dynamic workflow orchestration and scheduling mechanisms for complex working conditions. This makes it difficult to achieve unified management of data access, processing, retrieval, and inference processes, resulting in insufficient system flexibility and scalability. In terms of risk assessment methods, traditional statistical or machine learning models are often "black box" structures, only outputting prediction results. They lack a clear risk scoring construction logic and a multi-dimensional feature fusion mechanism, making it difficult to reflect the coupling effects of multiple factors such as surrounding rock deformation, stress, and energy release, leading to weak interpretability of the prediction results. Furthermore, current technologies for utilizing historical cases and engineering standards largely rely on manual experience, lacking a systematic retrieval and fusion mechanism. This makes it difficult to effectively combine historical experience with real-time data, hindering the formation of evidence-based reasoning processes and thus failing to support refined risk assessment and decision-making under complex operating conditions. Therefore, achieving unified expression and consistency control of multi-source data, constructing schedulable intelligent processing flows, establishing risk scoring models with engineering significance, and combining historical cases and regulatory rules for reasoning correction have become critical technical issues that urgently need to be addressed. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is: existing methods for predicting the stability of coal mine surrounding rock have the following problems: it is difficult to uniformly access and integrate multi-source heterogeneous data; the data processing flow lacks dynamic arrangement and scheduling mechanisms, resulting in fragmented analysis processes; the risk assessment model has insufficient interpretability and is difficult to reflect the coupling effect of multiple physical fields; and how to achieve intelligent prediction and decision support for the stability of surrounding rock based on the fusion of structured features, historical cases, and normative rules.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: an intelligent prediction method for the stability of surrounding rock in underground coal mine roadways, comprising: collecting real-time data streams related to the working conditions of the surrounding rock in the roadway; attaching a unified timestamp and collection location identifier to each data point to form a multi-source data set with temporal consistency and spatial correlation; simultaneously performing data consistency control and cross-measuring point correlation analysis; uploading the multi-source data set to an intelligent agent through a data interface; using the Dify workflow orchestration engine for node-based orchestration and dynamic scheduling execution; constructing a structured feature set with a unified time step and spatial correlation; constructing a surrounding rock stability risk scoring model based on the structured feature set; and combining historical cases and normative rules obtained through retrieval as the basis for reasoning input into the DeepSeek large language model for reasoning correction; and outputting the surrounding rock stability level and prediction results.

[0007] As a preferred embodiment of the intelligent prediction method for the stability of surrounding rock in underground coal mine roadways according to the present invention, the formation of a multi-source data set with temporal consistency and spatial correlation includes: deploying a monitoring network in the underground coal mine roadways consisting of displacement monitoring sensors, stress monitoring sensors, microseismic monitoring equipment, and seepage monitoring devices; displacement monitoring is used to obtain the deformation and convergence trend of the surrounding rock, stress monitoring is used to obtain the stress changes of anchor bolts and surrounding rock, microseismic monitoring is used to record the energy release and fracturing activity inside the surrounding rock, seepage monitoring is used to reflect changes in groundwater pressure, and environmental parameters are used to describe changes in temperature and gas conditions; during the data acquisition process, various sensors continuously output data according to their respective sampling mechanisms and uniformly convert them into data records with time stamps and spatial location descriptions; spatial identifiers are generated based on roadway mileage, cross-section number, and measuring point number, and combined with time stamps to form unique data identifiers; preliminary consistency verification is performed on the data at the acquisition end, and when duplicate data at the same time and location is detected, deduplication is performed by retaining the first arriving data; the processed data enters the buffer in chronological order and is input to the Dify workflow execution environment in streaming form.

[0008] As a preferred embodiment of the intelligent prediction method for the stability of surrounding rock in underground coal mine roadways according to the present invention, the data consistency control and cross-measuring point correlation analysis include: in the data access node of the workflow, uniformly parsing the input data and establishing a data index structure based on time markers and spatial identifiers; matching different sensor data through the index, associating data from different measuring points within the same time period to form a multi-dimensional data combination; in the data processing node, sorting the data according to the time window and identifying abnormal fluctuations; when a certain data shows a sudden change relative to the preceding and following time periods and is inconsistent with the change trend of adjacent measuring points, it is determined as abnormal data and removed; after removing abnormal data, smoothing and filling is performed based on the change trend of the preceding and following data; at the same time, by comparing the data differences between adjacent measuring points, spatial change information is extracted to reflect the local uneven deformation characteristics of the surrounding rock.

[0009] As a preferred embodiment of the intelligent prediction method for the stability of surrounding rock in underground coal mine roadways according to the present invention, the node-based orchestration and dynamic scheduling execution includes: constructing a Dify-based workflow structure in the intelligent agent, dividing the entire processing flow into data access nodes, condition judgment nodes, document parsing nodes, data processing nodes, database retrieval nodes, and network request nodes, and defining the execution order between nodes through process configuration; after receiving real-time data streams at the start node, the system enters the condition judgment node; when a document input is detected, the document parsing node is triggered to parse the file; otherwise, the system directly enters the data processing node; the data processing node performs unified format conversion on multi-source data and outputs standardized data; historical case data is obtained through the database retrieval node, and external specifications or knowledge information is supplemented through the network request node; each node executes automatically according to the workflow order, and the result is passed to the next node after the node completes its execution; when a node fails to execute, a retry or switch to a backup path is triggered through the workflow control mechanism.

[0010] As a preferred embodiment of the intelligent prediction method for the stability of surrounding rock in underground coal mine roadways according to the present invention, the structured feature set includes: aligning different types of monitoring data according to a unified time scale during the data processing stage of the workflow, so that each type of data has a corresponding value at the same time node; extracting deformation increment and trend features from displacement data, extracting stress change features from stress data, extracting energy release intensity and frequency features from microseismic data, and extracting pressure change features from seepage data; performing combined analysis on data from the same cross-section and adjacent measuring points based on spatial location to extract spatial difference features; and simultaneously performing trend analysis on time series data to extract change rate and fluctuation features.

[0011] As a preferred embodiment of the intelligent prediction method for the stability of surrounding rock in underground coal mine roadways described in this invention, the construction of the surrounding rock stability risk scoring model includes: inputting features from the structured feature set as different dimensions and setting corresponding weights for each dimension; during the calculation process, normalizing various features and then performing comprehensive weighted calculation to obtain a comprehensive risk score reflecting the current stability state of the surrounding rock; deformation features are used to reflect the degree of surrounding rock deformation, stress features are used to reflect the degree of stress concentration, and microseismic features are used to reflect the intensity of potential destructive activities; after the model calculation is completed, the obtained comprehensive risk score is used as the basic assessment result and transmitted to the large language model inference node.

[0012] As a preferred embodiment of the intelligent prediction method for the stability of surrounding rock in underground coal mine roadways described in this invention, the output of the surrounding rock stability level and prediction results includes: in the inference node of the workflow, the basic assessment results are fused with the current structured features and retrieved historical cases and normative rules and input into the DeepSeek large language model, which performs semantic correction and interpretation of the risk score and outputs the final stability level and key risk factors; in the early warning judgment node, the stability level and its changing trend are judged, and an early warning event is triggered when the preset level is reached or a continuous deterioration trend occurs; the generated early warning information includes the risk level, risk source and disposal suggestions, and is sent to the mine dispatch room terminal through the push node in the workflow.

[0013] As a preferred embodiment of the intelligent prediction system for the stability of surrounding rock in underground coal mine roadways according to the present invention, the system includes: a data acquisition module, a data upload module, and a prediction module. The data acquisition module collects real-time data streams related to the working conditions of the surrounding rock in the roadway, and adds a unified timestamp and acquisition location identifier to each data point to form a multi-source data set with temporal consistency and spatial correlation, while simultaneously performing data consistency control and cross-measuring point correlation analysis. The data upload module uploads the multi-source data set to the intelligent agent through a data interface, where it is arranged and dynamically scheduled by the Dify workflow orchestration engine to construct a structured feature set with a unified time step and spatial correlation. The prediction module constructs a surrounding rock stability risk scoring model based on the structured feature set, and combines retrieved historical cases and standard rules as the basis for reasoning, inputting them into the DeepSeek large language model for reasoning correction, and outputting the surrounding rock stability level and prediction results.

[0014] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement a method for intelligent prediction of the stability of surrounding rock in underground coal mine roadways.

[0015] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of an intelligent prediction method for the stability of surrounding rock in underground coal mine roadways.

[0016] The beneficial effects of this invention are as follows: The intelligent prediction method for the stability of surrounding rock in underground coal mine roadways provided by this invention introduces a unified timestamp and spatial identifier during the data acquisition stage, and combines it with joint judgment rules for temporal and spatial consistency. This achieves highly reliable input of multi-source heterogeneous monitoring data, effectively solving the problems of data dispersion, large noise interference, and difficulty in unified modeling in traditional methods. By introducing the Dify workflow orchestration engine, data access, document parsing, knowledge retrieval, and feature construction are organized and dynamically scheduled in a node-based manner, realizing the automation and structuring of the data processing flow. This allows multi-physics monitoring information to form a computable structured feature set under a unified time step and spatial correlation. On this basis, a risk scoring model integrating multi-dimensional information such as deformation, force, and energy is further constructed. Combined with historical cases and standard rules, semantic reasoning correction is performed through a large language model, so that the prediction results not only have quantitative evaluation capabilities but also interpretability and engineering decision-making guidance significance. Thus, this invention achieves end-to-end closed-loop processing from raw monitoring data to stability prediction and early warning output, improving the accuracy, reliability, and engineering applicability of surrounding rock stability prediction. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is an overall flowchart of an intelligent prediction method for the stability of surrounding rock in underground coal mine roadways, provided in Embodiment 1 of the present invention.

[0019] Figure 2 The Dify workflow diagram is provided for an intelligent prediction method for the stability of surrounding rock in underground coal mine roadways according to Embodiment 1 of the present invention.

[0020] Figure 3 This is a schematic diagram of a computer device for an intelligent prediction method of surrounding rock stability in underground coal mine roadways, as provided in Embodiment 3 of the present invention. Detailed Implementation

[0021] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0022] Example 1, referring to Figures 1-2 As an embodiment of the present invention, a method for intelligent prediction of the stability of surrounding rock in underground coal mine roadways is provided, comprising:

[0023] S1: Collect real-time data streams related to the working conditions of the surrounding rock in the roadway, and add a unified timestamp and collection location identifier to each data to form a multi-source data set with temporal consistency and spatial correlation, while performing data consistency control and cross-measuring point correlation analysis.

[0024] Furthermore, forming a multi-source data set with temporal consistency and spatial correlation involves deploying a monitoring network in underground coal mine roadways, consisting of displacement monitoring sensors, stress monitoring sensors, microseismic monitoring equipment, and seepage monitoring devices. Displacement monitoring is used to acquire the deformation and convergence trend of the surrounding rock; stress monitoring is used to acquire the stress changes of anchor bolts and surrounding rock; microseismic monitoring is used to record the energy release and fracturing activity within the surrounding rock; seepage monitoring is used to reflect changes in groundwater pressure; and environmental parameters are used to describe changes in temperature and gas conditions. During data acquisition, various sensors continuously output data according to their respective sampling mechanisms, and the data is uniformly converted into data records with time stamps and spatial location descriptions. Spatial identifiers are generated based on roadway mileage, cross-section number, and measuring point number, and combined with time stamps to form unique data identifiers. At the acquisition end, preliminary consistency checks are performed on the data. When duplicate data at the same time and location is detected, deduplication is performed by retaining the first arriving data. The processed data enters the buffer in chronological order and is then streamed into the Dify workflow execution environment.

[0025] It should also be noted that a preferred scheme for forming a multi-source data set with temporal consistency and spatial correlation specifically includes deploying a monitoring network along the roadway direction and key sections in the coal mine roadway to be monitored. This monitoring network includes at least displacement monitoring sensors, stress monitoring sensors, roof delamination monitoring devices, surrounding rock strain sensors, microseismic or acoustic emission monitoring equipment, pore pressure or seepage monitoring devices, and environmental parameter monitoring devices. Displacement sensors are used to reflect the convergence of the surrounding rock, the approach of the roof and floor, or the deformation trend of the sidewalls; stress sensors are used to reflect changes in the stress on anchor bolts, anchor cables, and local surrounding rock; roof delamination monitoring devices are used to reflect roof separation and signs of instability; surrounding rock strain sensors are used to reflect the accumulation of local deformation in the surrounding rock; microseismic or acoustic emission monitoring equipment is used to record the propagation of internal fractures and energy release activities in the surrounding rock; pore pressure or seepage monitoring devices are used to reflect groundwater pressure or hydraulic disturbances; and environmental parameter monitoring devices are used to record changes in external conditions such as temperature, humidity, and gas concentration. For construction disturbance parameters, information such as blasting time, explosive charge, tunneling advance, and support operation time can also be collected and incorporated into the same data recording system along with the aforementioned monitoring values.

[0026] During data collection, each original monitoring record was recorded as follows:

[0027]

[0028] in, Indicates the first Class of monitoring data, in the first The measuring point, the first One original data record corresponds to one sampling record. This indicates the data type number, used to distinguish data from different sources such as displacement, stress, roof delamination, surrounding rock strain, microseismic activity, seepage, environmental parameters, and construction disturbance parameters; This indicates the measuring point number, which corresponds to a fixed installation location within the roadway. This indicates the first measurement point under the same data type. One original sampling record; This indicates the monitoring value of this record; This represents the unified timestamp at which the record was generated; The spatial location identifier of the measuring point is represented as follows:

[0029]

[0030] in, This indicates the mileage location of the measuring point in the tunnel. This indicates the section number to which the measuring point belongs. This indicates the measurement point number within the corresponding cross-section. To eliminate record confusion caused by different sensors, different cache batches, or different upload latencies at the data acquisition source, a unique identifier is generated for each record in this embodiment, represented as:

[0031]

[0032] in, Indicates the first Class data in the first The first measuring point Each record's unique identifier consists of three parts: data type, timestamp, and spatial location. This design aims to ensure that even if different devices use different upload paths or upload orders, records of the same data type, uploaded at the same time, and in the same spatial location can be identified as duplicate records from the same source before entering the subsequent system. To handle inconsistencies in upload order, this embodiment appends an arrival sequence number to each arriving data record at the acquisition end or upload interface. And perform duplicate resolution according to the following rules: when there are two records that have the same unique identifier, i.e. and When the arrival sequence number is smaller, keep the record with the smaller arrival sequence number and delete the rest. Indicates the first Class data in the first The first measuring point A unique identifier for each record. Here... This indicates the order in which the record arrives at the acquisition interface or buffer; a smaller value indicates an earlier arrival. This "unique identifier + first-come-first-served" consistency verification rule at the acquisition end avoids data bloat and false anomalies caused by network retransmissions, edge buffer retransmissions, or instantaneous duplicate uploads from devices. Because various sensors use different sampling methods, to ensure that subsequent cross-type data can be compared on the same time base, this embodiment does not directly send all raw data to the subsequent intelligent agent one by one after acquisition, but rather organizes it according to a unified time window. The start time of the window is set to... Window length is Then the first Time window Represented as:

[0033]

[0034] in, The time window number, To standardize the time window length, it can be set to 1 minute, 5 minutes, or other fixed durations as needed on site. All that meets the requirements... The original records are all classified under the first... This is done within a time window. In this way, raw data with different frequencies and triggering mechanisms are first organized into a unified time slice, so that the states of different types of data can be compared at the same analysis time.

[0035] For entering the same time window The original record, in this embodiment, selects a window representative value based on the data type, denoted as... Its definition is:

[0036]

[0037] in, Indicates the first Class data in the first The measuring point, the first Representative values ​​within a time window; Representation and data type Corresponding window aggregation rules. For data reflecting the maximum deformation risk, such as displacement and roof delamination, the maximum value within the window can be taken as the representative value; for continuous and stable data such as anchor stress, surrounding rock strain, pore pressure, and environmental parameters, the average or final value within the window can be taken; for microseismic or acoustic emission event data, the number of events, cumulative energy, or maximum energy within the window can be converted into the representative value of that window; for construction disturbance parameters, the corresponding construction status or construction intensity characterization quantity can be extracted based on whether blasting, tunneling, or support operations have occurred within the window. Thus, In fact, the original multi-frequency data has been converted into a unified time expression result in window units, which is the direct input for subsequent anomaly identification and cross-measurement point correlation analysis.

[0038] It should be noted that data consistency control and cross-measuring point correlation analysis include: in the data access node of the workflow, uniformly parsing the input data and establishing a data index structure based on time stamps and spatial identifiers; matching data from different sensors through the index, associating data from different measuring points within the same time period to form a multi-dimensional data combination; in the data processing node, sorting the data by time window and identifying abnormal fluctuations; when a certain data point shows a sudden change relative to the preceding and following time periods and is inconsistent with the change trend of adjacent measuring points, it is judged as abnormal data and removed; after removing abnormal data, smoothing and filling in the data based on the change trend of the preceding and following data; and simultaneously, by comparing the data differences between adjacent measuring points, extracting spatial change information to reflect the local uneven deformation characteristics of the surrounding rock.

[0039] It should also be noted that a preferred scheme for data consistency control and cross-measurement point correlation analysis specifically includes, in this embodiment, establishing an index structure in the data access node of the Dify workflow after obtaining the window representative value, to achieve cross-type and cross-measurement point data correlation within the same time window. For any window Construct the data set for this window:

[0040]

[0041] in, Indicates the first A collection of all deduplicated, window-representative data within a given time window. Each element in the collection also carries a data type number. Measurement point number Window representative value and spatial location Based on this set, the system can use "combinations of different sensors within the same window" or "combinations of adjacent measuring points within the same window" as index objects to form multidimensional data combinations, providing a foundation for data consistency control and spatial correlation analysis.

[0042] Unlike conventional methods that rely solely on sudden changes in time at a single measurement point to identify anomalies, this embodiment employs a combined rule of "temporal consistency + spatial consistency" to perform data consistency control. First, a time reference value is established for the same data type at the same measurement point:

[0043]

[0044] in, Indicates the first Class data in the first The measuring point, the first The time base value corresponding to each time window Indicates the historical step index, Indicates the first Class data in the first The measuring point, the first Representative values ​​within a time window; This indicates the number of windows for backtracking this type of data. Different values ​​can be set depending on the data type. For example, displacement and stress data can use shorter backtracking windows, while environmental and seepage data can use longer backtracking windows.

[0045] Based on this time baseline, the deviation of the current window from the previous trend can be calculated:

[0046]

[0047] in, It represents the time deviation, used to characterize the degree of abrupt change in the current monitoring value relative to the historical trend of the monitoring point itself.

[0048] Secondly, to determine whether the abrupt change is consistent with the surrounding rock condition, this embodiment establishes spatial reference values ​​between adjacent measuring points for the same data type. Let... Indicates the measuring point Given the set of neighboring measurement points, we have:

[0049]

[0050] in, Indicates the first Class data in the first The measuring point, the first Spatial reference value for a time window, Indicates the first Class data in the first The measuring point, the first Representative values ​​within a time window; Represents the set of neighboring measurement points A certain measuring point in; This indicates the number of neighboring measurement points.

[0051] Based on this spatial reference value, the spatial deviation of the current measuring point from the group state of adjacent measuring points can be calculated. , represented as:

[0052]

[0053] in, It is used to characterize whether the current measuring point deviates from the overall trend of change of the same or adjacent cross sections.

[0054] In this embodiment, the window's representative value is only considered abnormal when both "significant time deviation" and "significant spatial deviation" conditions are met simultaneously, to avoid mistakenly deleting real disaster precursors as noise. The abnormality determination variable is denoted as... The rules are written as follows:

[0055]

[0056] in, Indicates the first Class data in the first The measuring point, the first One time window was judged to be abnormal; Indicates no abnormality; Indicates the first Time deviation threshold for class data; Indicates the first Spatial deviation thresholds for different data types. A categorized threshold is used here instead of a uniform threshold because the fluctuation scales of data such as displacement, stress, microseismic activity, and seepage have different engineering implications. For example, microseismic energy can allow for a short-term surge after localized blasting, while a rapid and inconsistent increase in roof delamination at adjacent measuring points is more likely to correspond to signs of local instability. Therefore, through... and The categorization settings allow anomaly identification rules to be tailored to the coal mine surrounding rock monitoring scenario, rather than remaining at the level of general signal cleaning.

[0057] when If the value represented by the window deviates from both its own historical trend and the trend of neighboring monitoring points, it should be removed from the valid monitoring sequence, and the continuous monitoring data should be smoothed and completed. The completed value is denoted as... The calculation is as follows:

[0058]

[0059] in, This represents the effective window value after consistency control. Indicates the first Class data in the first The measuring point, the first Representative values ​​within a time window Indicates the first Class data in the first The measuring point, the first For continuous data such as displacement, stress, roof delamination, strain, seepage, and environmental parameters, the representative values ​​within a time window can be smoothly supplemented using the representative values ​​from preceding and following windows. For microseismic or acoustic emission event data, direct interpolation is not necessary; instead, the abnormal window is marked as an "event window requiring manual review" or updated using statistical smoothing of events from adjacent windows, thus avoiding completely erasing real sudden events. This setup aims to ensure that different types of data are processed using methods adapted to their physical properties during quality control, maintaining sequential continuity without compromising engineering interpretability. After anomaly identification and supplementation, this embodiment further extracts the spatial difference across measuring points to describe local uneven deformation of the surrounding rock. For the first... Class data at measurement points Spatial difference characteristics at a location are defined as follows:

[0060]

[0061] in, This represents the deviation of the measurement point from the group state of neighboring measurement points after consistency control. If... A consistently positive and gradually increasing value indicates that the location of the measuring point may experience localized stress concentration, intensified delamination, or uneven convergence; a consistently negative value indicates that the point may be in a state of release or regression relative to its surroundings. This spatial difference is not constructed temporarily in subsequent steps, but is formed along with the effective window value in stage S1, allowing it to be directly invoked when constructing the structured feature set in subsequent workflows. As a spatial correlation input, without having to retrace the original sampling stream.

[0062] Finally, after the process of "collection – unified identification – time window organization – window representative value extraction – temporal consistency determination – spatial consistency determination – anomaly removal and smooth completion – spatial difference extraction", the first... The effective multi-source data set corresponding to each time window is denoted as:

[0063]

[0064] in, This represents a multi-source data set that has undergone time and space consistency processing; each element contains a data type. Measurement point number Effective window value Spatial differences Spatial location identifiers and time window This set, as the output of step S1, is sent to subsequent nodes in the Dify workflow execution environment for the next step of structured feature construction, risk scoring model calculation, and large language model inference correction. A monitoring section can be set up every 20 meters along the return airway of a coal mine working face. Each section has measuring points at the roof centerline, left sidewall, right sidewall, and floor. Displacement monitoring sensors record the surrounding rock convergence, stress sensors record the anchor bolt axial force, micro-vibration equipment records event energy, and seepage sensors record pore pressure changes.

[0065] Set a unified time window After minutes, all sensor data within any window are first aggregated into Subsequently, the time deviation Spatial deviation The joint judgment is abnormal.

[0066] If the separation value of a certain top plate measuring point suddenly increases significantly in the current window compared to the previous few windows, and no similar change is observed in adjacent measuring points on the same cross section, then the value is marked as abnormal and completed; if multiple adjacent measuring points show synchronous uplift, it will not be simply deleted, but will be retained as a valid risk indicator for subsequent analysis.

[0067] It should also be noted that this study addresses the challenge of constructing a unified spatiotemporal data foundation directly applicable to intelligent analysis in the multi-source, heterogeneous monitoring environment of underground coal mine roadways, characterized by different sampling mechanisms and spatial distributions. Coal mine surrounding rock monitoring data simultaneously includes various data such as displacement, stress, roof delamination, microseismic activity, seepage, environmental parameters, and construction disturbances. These data differ in sampling frequency, triggering mode, spatial location, and engineering meaning. Simply using traditional methods for independent collection or simple aggregation would be insufficient to support subsequent unified analysis and would fail to guarantee the reliability of the results. By attaching a unified timestamp and collection location identifier to each data point, establishing a unique data identifier, and combining time window organization, window representative value extraction, duplicate record elimination, and a joint judgment mechanism of "temporal consistency + spatial consistency," a unified expression and quality control of multi-source data in both temporal and spatial dimensions is achieved. This not only improves the authenticity and continuity of the input data but also avoids mistakenly deleting local instability precursors as noise.

[0068] S2: Upload multi-source data sets to the intelligent agent through the data interface, and the Dify workflow orchestration engine performs node-based orchestration and dynamic scheduling execution to build a structured feature set with unified time steps and spatial relationships.

[0069] Furthermore, node-based orchestration and dynamic scheduling execution include building a Dify-based workflow structure within the agent, dividing the entire processing flow into data access nodes, condition judgment nodes, document parsing nodes, data processing nodes, database retrieval nodes, and network request nodes, and defining the execution order between nodes through process configuration; after receiving the real-time data stream at the start node, it enters the condition judgment node. When a document input is detected, the document parsing node is triggered to parse the file; otherwise, it directly enters the data processing node; the data processing node performs unified format conversion on multi-source data and outputs standardized data; historical case data is obtained through the database retrieval node, and external specifications or knowledge information is supplemented through the network request node; each node executes automatically according to the workflow order, and the result is passed to the next node after the node completes its execution; when a node fails to execute, a retry or switch to a backup path is triggered through the workflow control mechanism.

[0070] It should also be noted that a preferred scheme for node-based orchestration and dynamic scheduling execution specifically includes constructing a Dify-based workflow structure within the agent, dividing the entire processing into the following node sequence: data access node, condition judgment node, document parsing node, data organization node, data processing node, database retrieval node, network request node, and feature construction node; the nodes are connected according to a preset process to form a directed execution chain. For the first... Data in a time window The workflow execution process can be represented as:

[0071]

[0072] in: Indicates the data access node. This represents a conditional decision node. This represents the set of nodes related to data processing and organization. Represents the feature-based construction node.

[0073] At the data access node In the input set Perform parsing and extract data types. Measurement point number Valid value Spatial differences and location markers And establish an internal index structure to enable different types of data to be accessed uniformly within the same time window. In the condition judgment node... In this process, it is determined whether the input data contains external documents or additional information. Let the input flag variable be... When there is document input ,otherwise .when When the document parsing node is triggered, the project description, construction records, or exception descriptions in the document are extracted into text information; when At that time, it directly enters the data processing node.

[0074] refer to Figure 2 This includes detailed functionality of each module in Dify:

[0075] Start: In the UI, provide input boxes to collect user-provided information, including text, documents, and charts.

[0076] Document formats include TXT, MD, MDX, MARKDOWN, PDF, HTMLXLSX, XLS, DOC, DOCX, CSV, EML, MSG, PPTX, PPT, XML, EPUB, etc., and image formats include JPG, JPEG, PNG, GIF, WEBP, and SVG, etc.

[0077] Conditional branch: Determine if the user has entered a document or file. If so, extract the document; otherwise, skip document extraction and proceed to the collection and organization node.

[0078] Document Extractor: Extracts data from user-provided documents.

[0079] Collection and organization: Use the DeepSeek large model to collect text, documents, and charts from the "Start" node and the "Document Extractor" node. Organize the content format so that the large model can understand it.

[0080] Database: A historical database formed from field data.

[0081] HTTP Request: Web Search Request. Based on workflow requirements, a web search is performed to fill in any gaps in the data.

[0082] Comprehensive Analysis and Reasoning: Utilizing the Deepseek large-scale model for comprehensive analysis and reasoning. Data collected from databases and online searches is comprehensively analyzed and reasoned to arrive at the criteria for judging the stability of the surrounding rock in the tunnel.

[0083] Parameter Extractor: Extracts the content from the "Comprehensive Analysis and Reasoning" node. It divides the content of the previous node into names and task descriptions for multiple steps, which are then passed to the next "Iteration" node for execution.

[0084] Iteration: This includes code execution and DeepSeek large model inference analysis. The "Code Execution" node compiles the code based on the name and task description of the previous node, allowing the DeepSeek large model to execute it; the "DeepSeek Large Model Inference Analysis" node performs inference analysis on the previous node; the number of iterations is automatically determined based on the amount of tasks, completing the task iteratively.

[0085] Template conversion: Format conversion and adjustment, such as "line break, bold", convert according to preset format for easy reading.

[0086] Intelligent prediction: Based on the reasoning and analysis of the above iterative content, the current status of roadway stability is qualitatively and quantitatively determined, and intelligent prediction is made.

[0087] Reply: Reply according to the preset output format.

[0088] It should be noted that the structured feature set includes, during the data processing stage of the workflow, aligning different types of monitoring data according to a unified time scale so that each type of data has a corresponding value at the same time node; extracting deformation increment and trend features from displacement data, extracting stress change features from stress data, extracting energy release intensity and frequency features from microseismic data, and extracting pressure change features from seepage data; performing combined analysis on data from the same cross section and adjacent measuring points based on spatial location to extract spatial difference features; and simultaneously performing trend analysis on time series data to extract change rate and fluctuation features.

[0089] It should also be noted that a preferred approach to constructing a structured feature set with a unified time step and spatial correlation specifically includes, at the data processing node, performing a unified format conversion on data from different sources. This process maps the input set to a standardized data set. , represented as:

[0090]

[0091] The time window identifier has been removed. However, retain the time series index. .

[0092] To enable joint analysis of different types of data under the same time benchmark, this embodiment performs the following in the data processing node: Perform time series alignment processing. For any measurement point and data types Constructing time series , represented as:

[0093]

[0094] Based on this sequence, the time variation characteristics can be further calculated, expressed as:

[0095]

[0096] in Indicates the first Class data at measurement points The change in the value is used to reflect short-term trends. Simultaneously, a local average value is introduced to reflect the smoothness of the change. , represented as:

[0097]

[0098] in This indicates the length of the sliding window. This value is used to describe the average state of the measurement point over a recent time period.

[0099] Based on spatial difference characteristics This embodiment further constructs a cross-sectional spatial consistency index. For the same cross-section... Suppose that the cross-section contains a set of measuring points. Then the cross-sectional average value is defined. , represented as:

[0100]

[0101] Based on this definition, the characteristics of cross-sectional deviation are:

[0102]

[0103] in Indicates the measuring point The degree of deviation relative to its cross section is used to reflect local anomalies or uneven deformation.

[0104] In this embodiment, different types of data are no longer used independently, but are combined to form multiphysics features. For the same measurement point... Construct feature vectors:

[0105]

[0106] in: Indicates characteristics of change over time. Indicates trend characteristics, Indicates spatial neighborhood differences, This indicates cross-sectional consistency deviation, and the feature vector covers multi-dimensional information such as deformation, stress, energy release, and hydraulics. (In the feature construction node...) In the middle, all measuring points are in the first... The feature vectors from each time window are summarized to form a structured feature set:

[0107]

[0108] in, Indicates the first The complete structured input within a time window contains both numerical features and retains spatial location information.

[0109] It should also be noted that, in order to address the challenge of achieving unified orchestration and dynamic execution of data processing, external information access, and feature construction processes in the continuous, complex, and multi-input-source scenario of coal mine surrounding rock monitoring, thus forming a truly implementable intelligent analysis chain, the common practice in existing technologies is to implement data cleaning, feature extraction, knowledge supplementation, and model invocation independently. While the individual functional modules themselves are not unfamiliar, in the context of coal mine surrounding rock monitoring, data sources include not only real-time sensor data but also construction records, documents, standard texts, and historical cases. This requires the system to not only process "data" but also "heterogeneous information input" and dynamically determine the processing path based on the input type. By introducing the Dify workflow orchestration engine, data access nodes, condition judgment nodes, document parsing nodes, data processing nodes, database retrieval nodes, network request nodes, and feature construction nodes are integrated into a unified workflow execution system, solving the problems of fragmented processing chains, fixed processes, and difficulty in inter-node collaboration under complex input conditions. Building upon this foundation, a structured feature set with a unified time step and spatial correlation is constructed. This transforms subsequent modeling inputs from simple monitoring values ​​into structured expressions covering temporal variations, spatial differences, and multi-physics coupling information. This enhances the system's automated execution capabilities, process stability, and scalability, and for the first time truly integrates multi-source data processing with knowledge supplementation, engineering data analysis, and feature construction into a unified process.

[0110] S3: Construct a surrounding rock stability risk scoring model based on a structured feature set, and combine it with historical cases and normative rules obtained from retrieval as the basis for reasoning. Input the data into the DeepSeek large language model for reasoning correction, and output the surrounding rock stability level and prediction results.

[0111] Furthermore, the construction of the surrounding rock stability risk scoring model includes: taking the features in the structured feature set as inputs for different dimensions and setting corresponding weights for each dimension; during the calculation process, normalizing the various features and then performing a comprehensive weighted calculation to obtain a comprehensive risk score reflecting the current stability state of the surrounding rock; deformation features are used to reflect the degree of deformation of the surrounding rock, stress features are used to reflect the degree of stress concentration, and microseismic features are used to reflect the intensity of potential destructive activities; after the model calculation is completed, the obtained comprehensive risk score is used as the basic assessment result and passed to the inference node of the large language model.

[0112] It should also be noted that a preferred scheme for constructing a surrounding rock stability risk scoring model based on a structured feature set specifically includes, in order to achieve quantitative assessment of surrounding rock stability, dividing features with different physical meanings into three core risk dimensions: deformation dimension (corresponding to displacement, delamination, strain, etc.), stress dimension (corresponding to anchor bolt axial force, surrounding rock stress, etc.), and energy dimension (corresponding to microseismic or acoustic emission activities). The comprehensive characterization quantity for each dimension is defined as follows:

[0113] The deformation risk characterization quantity is jointly determined by displacement increment, delamination change, and spatial differences;

[0114] The quantity representing stress risk is determined by stress changes and their spatial concentration.

[0115] It represents the energy risk characterization quantity, which is determined by the frequency of microseismic events and the intensity of energy release.

[0116] During the calculation, all types of features are first subjected to a uniform scaling process to ensure they fall within a comparable range. For any feature quantity... Its normalization result is denoted as:

[0117]

[0118] in and These represent the minimum and maximum values ​​of this type of data within the historical monitoring range, respectively, to ensure a uniform scale across different physical quantities. Subsequently, a weighted aggregation is performed on each risk dimension, such as the deformation dimension:

[0119]

[0120] in This represents a collection of data types that belong to the variant class. This represents the corresponding feature weight, whose values ​​are set based on historical engineering experience or specifications; the force dimension and energy dimension are calculated in the same way. After obtaining the risk quantities for each dimension, a comprehensive risk score is constructed:

[0121]

[0122] in Indicates the measuring point In the time window The overall risk score below; For the dimension weight coefficients, satisfying Its value can be adjusted according to the geological conditions of the mining area or the support method. For example, the weight of the deformation dimension can be increased in soft rock tunnels, and the weight of the energy dimension can be increased in strong disturbance conditions. To avoid the score only reflecting the instantaneous state, this embodiment further introduces a trend enhancement factor. Define the risk change rate:

[0123]

[0124] in, Indicates the first Each measuring point in the current time window The change in risk below, Indicates the measuring point In the time window The overall risk score is as follows.

[0125] It should be noted that the output of the surrounding rock stability level and prediction results includes the following steps: In the inference node of the workflow, the basic assessment results are integrated with the current structured features and retrieved historical cases and normative rules, and input into the DeepSeek large language model. The DeepSeek large language model performs semantic correction and interpretation of the risk score and outputs the final stability level and key risk factors. In the early warning judgment node, the stability level and its changing trend are judged. When the preset level is reached or a continuous deterioration trend occurs, an early warning event is triggered. The generated early warning information includes the risk level, risk source and disposal suggestions, and is sent to the mine dispatch room terminal through the push node in the workflow.

[0126] It should also be noted that a preferred scheme for outputting the surrounding rock stability level and prediction results specifically includes obtaining the basic risk score. Subsequently, this embodiment does not directly output the results, but instead performs further corrections through the retrieval and inference nodes in the workflow. First, in the database retrieval node, based on the current structured feature set... Retrieve a set of cases similar to the current working conditions from the historical case database. , denoted as:

[0127]

[0128] in, Indicates the first Each of the retrieved historical cases It includes historical monitoring characteristics, surrounding rock stability results, and remedial measures. Simultaneously, it obtains the currently applicable set of rules and regulations through network request nodes. , represented as:

[0129]

[0130] in, Indicates the current time window The first search result obtained Standard rules or engineering constraints, rules This includes threshold conditions, support requirements, and safety specifications. Subsequently, a unified input is constructed in the inference node:

[0131]

[0132] in, This represents the comprehensive inference input set fed into the large language model. This input includes numerical scores, feature information, historical evidence, and canonical constraints. This input is then passed to the DeepSeek large language model, which performs semantic inference and outputs the corrected result.

[0133]

[0134] in: Indicates the final stability level (e.g., I to V). Explanation of key risk factors; This indicates the recommended handling measures. Unlike directly outputting scores, the role of the large language model in this embodiment is to provide "interpretive correction" to the scores. For example, when a score is high but inconsistent with historical cases, the model can adjust its judgment based on evidence or point out the reason for the anomaly.

[0135] After the inference node outputs its results, it proceeds to the early warning judgment node. This is based on the stability level. and risk change trends To make an early warning judgment:

[0136] An alert is triggered when any of the following conditions are met: Reaching the preset high-risk level; Exceeding the preset threshold; It remains positive and exceeds the change threshold.

[0137] An alert message is generated after the alert is triggered. , represented as:

[0138]

[0139] This includes risk level, risk score, key factors, and recommended measures.

[0140] The warning message is sent to the mine dispatch room terminal via a push node in the workflow, along with the input data. Search results , The inference output is stored together for subsequent analysis and model optimization.

[0141] In this embodiment, after obtaining the trend-corrected risk score... Next, the risk scores are processed using a level mapping method. Specifically, based on a preset scoring range, the risk scores are divided into multiple stability level ranges, and a level mapping function is defined:

[0142]

[0143] in, This represents the initial stability level obtained based on the scoring model. This represents the mapping rule from rating to level. Preferably, the rating interval is divided as follows: when... When, it is classified as Level I (stable); when When, it is judged as Level II (minor deformation); when When, it is judged as Level III (significant deformation); when When the condition is deemed unfavorable, it is classified as Level IV (instability warning); in some implementations, this can be further extended to Level V to accommodate extremely unstable conditions. To make the level determination more consistent with engineering practice, this embodiment further introduces a verification mechanism based on the surrounding rock deformation index. For the deformation index corresponding to the roof settlement, its window representative value is set to... Then define the engineering grade reference function:

[0144]

[0145] in, This indicates the engineering experience level based on the amount of roof settlement. Let be the mapping function from deformation to stability level, and its correspondence is: when It is classified as Class I; It was classified as Category II; when It was classified as Category III; It is classified as Category IV; It was classified as Category V. In practical applications, to avoid misjudgment based on a single indicator, the scoring level and the project level are consistently integrated to obtain the final level:

[0146]

[0147] in, This indicates the risk level fusion rule: when both risk levels match, the higher risk level is directly adopted; when they do not match, the higher risk level is prioritized as the final judgment result to ensure safety. In a specific example, when a measurement point calculates... Then, according to the scoring mapping rules, we get Simultaneously, if the roof settlement is 500mm, the corresponding engineering level is Class IV, ultimately determined as a Level IV instability warning. (The above levels...) As input, with a set of structured features Collection of historical cases and set of rules They are input into a large language model for semantic reasoning correction. Referring to Table 1, the stability level is usually divided into five categories based on the geological conditions of the surrounding rock, rock strength, and engineering practice requirements, and is ordered from high to low.

[0148] Table 1 Stability Level Table

[0149] Class I Extremely stable surrounding rock 10-50mm (average 30mm) The rock mass is intact, has high strength, poorly developed structural planes, and strong self-stabilizing ability. Category II Stable surrounding rock 50-100mm (average 75mm) The rock mass is relatively intact, has high strength, and good stability. Category III Medium-stability surrounding rock 100-400mm (average 250mm) The rock mass is relatively fractured and of medium strength, requiring appropriate support. Class IV Unstable surrounding rock 400-600mm (average 500mm) The rock mass is fractured, has low strength, and exhibits significant deformation, requiring enhanced support. Category V Extremely unstable surrounding rock 600-1800mm (average 1200mm) The rock mass is extremely fragmented, with very low strength and large deformation, requiring strong support.

[0150] Among them, Class III (moderately stable) and Class IV (unstable) are the most common levels that require reinforced support in engineering projects.

[0151] It should also be noted that, to address the challenge of transforming multi-source monitoring characteristics into stability prediction results that are quantifiable, interpretable, and consistent with engineering specifications and historical experience under complex surrounding rock conditions, traditional methods either rely on empirical thresholds and single-index judgments, making it difficult to handle multi-factor coupling scenarios; or they rely on machine learning black-box models, which, while potentially providing classification results, lack clear scoring logic and engineering interpretability, and cannot effectively embed historical cases and regulatory rules into the prediction process. Firstly, a surrounding rock stability risk scoring model is constructed based on a structured feature set, normalizing and weighting features such as deformation, stress, and energy dimensions to form a comprehensive risk score with clear physical meaning and engineering traceability. Subsequently, risk change rate, grade mapping, roof subsidence engineering verification, and historical case sets and regulatory rule sets are introduced as reasoning bases and input into the DeepSeek large language model for semantic correction. This achieves quantitative assessment and graded judgment of surrounding rock risk on the one hand, and enhances the rationality, interpretability, and engineering applicability of the prediction results through correction and interpretation under evidence constraints via the large language model on the other hand, while also directly outputting key risk factors and treatment recommendations.

[0152] Example 2, an embodiment of the present invention, provides an intelligent prediction system for the stability of surrounding rock in underground coal mine roadways, including a data acquisition module, a data upload module, and a prediction module.

[0153] The data acquisition module collects real-time data streams related to the surrounding rock conditions of the roadway, and adds a unified timestamp and collection location identifier to each data point to form a multi-source data set with temporal consistency and spatial correlation. It also performs data consistency control and cross-measuring point correlation analysis. The data upload module uploads the multi-source data set to the agent through a data interface. The Dify workflow orchestration engine performs node-based orchestration and dynamic scheduling execution to build a structured feature set with a unified time step and spatial correlation. The prediction module builds a surrounding rock stability risk scoring model based on the structured feature set, and combines historical cases and normative rules obtained from retrieval as the basis for reasoning input into the DeepSeek large language model for reasoning correction, and outputs the surrounding rock stability level and prediction results.

[0154] Example 3, referring to Figure 3This embodiment also provides a computer device applicable to the intelligent prediction method for the stability of surrounding rock in underground coal mine roadways, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the intelligent prediction method for the stability of surrounding rock in underground coal mine roadways as proposed in the above embodiment.

[0155] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0156] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the intelligent prediction method for the stability of surrounding rock in underground coal mine roadways as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

Claims

1. A method for intelligent prediction of the stability of surrounding rock in underground coal mine roadways, characterized in that, include: Real-time data streams related to the working conditions of the surrounding rock in the roadway are collected, and each data point is given a unified timestamp and a collection location identifier to form a multi-source data set with temporal consistency and spatial correlation. At the same time, data consistency control and cross-measuring point correlation analysis are performed. The multi-source data set is uploaded to the intelligent agent through the data interface, and the Dify workflow orchestration engine performs node-based orchestration and dynamic scheduling execution to build a structured feature set with unified time steps and spatial relationships. A surrounding rock stability risk scoring model is constructed based on a set of structured features. Historical cases and normative rules obtained from retrieval are used as the basis for reasoning and input into the DeepSeek large language model for reasoning correction. The model outputs the surrounding rock stability level and prediction results.

2. The intelligent prediction method for the stability of surrounding rock in underground coal mine roadways as described in claim 1, characterized in that: The formation of a multi-source data set with temporal consistency and spatial correlation includes, A monitoring network consisting of displacement monitoring sensors, stress monitoring sensors, microseismic monitoring equipment, and seepage monitoring devices is deployed in underground coal mine roadways; Displacement monitoring is used to obtain the deformation and convergence trend of the surrounding rock; stress monitoring is used to obtain the stress changes of the anchor bolt and the surrounding rock; microseismic monitoring is used to record the energy release and fracturing activity inside the surrounding rock; seepage monitoring is used to reflect the changes in groundwater pressure; and environmental parameters are used to describe changes in temperature and gas conditions. During the data acquisition process, various sensors continuously output data according to their respective sampling mechanisms, and the data is uniformly converted into data records with time stamps and spatial location descriptions. Spatial identifiers are generated based on tunnel mileage, cross-section number, and measuring point number, and combined with time markers to form unique data identifiers. At the data acquisition end, a preliminary consistency check is performed on the data. When duplicate data at the same time and location is detected, deduplication is performed by retaining the earlier data. The processed data enters the buffer in chronological order and is then fed into the Dify workflow execution environment in a streaming format.

3. The intelligent prediction method for the stability of surrounding rock in underground coal mine roadways as described in claim 2, characterized in that: The data consistency control and cross-measurement point correlation analysis include, In the data access node of the workflow, the input data is uniformly parsed and a data index structure is established based on time stamps and spatial identifiers; By matching data from different sensors using an index, data from different measuring points within the same time period can be associated to form a multi-dimensional data combination. In the data processing node, the data is sorted by time window and abnormal fluctuations are identified. When a certain data point changes abruptly relative to the preceding and following time periods and is inconsistent with the trend of changes of adjacent measuring points, it is judged as abnormal data and is removed. After removing outlier data, smooth data is added based on the trend of data changes before and after. Simultaneously, by comparing the data differences between adjacent measuring points, spatial variation information is extracted to reflect the local uneven deformation characteristics of the surrounding rock.

4. The intelligent prediction method for the stability of surrounding rock in underground coal mine roadways as described in claim 3, characterized in that: The node-based orchestration and dynamic scheduling execution includes... A Dify-based workflow structure is built in the intelligent agent, and the entire processing flow is divided into data access nodes, condition judgment nodes, document parsing nodes, data processing nodes, database retrieval nodes, and network request nodes. The execution order between nodes is defined through process configuration. After receiving the real-time data stream at the start node, the system enters the condition judgment node. When a document input is detected, the document parsing node is triggered to parse the file; otherwise, the system directly enters the data processing node. The data processing node performs unified format conversion on multi-source data and outputs standardized data; Historical case data is retrieved through database retrieval nodes, and external specifications or knowledge information is supplemented by network request nodes. Each node executes automatically according to the workflow sequence, and passes the result to the next node after the node has completed its execution; When a node fails to execute, a retry or switch to an alternative path is triggered through the workflow control mechanism.

5. The intelligent prediction method for the stability of surrounding rock in underground coal mine roadways as described in claim 4, characterized in that: The structured feature set includes, During the data processing phase of the workflow, different types of monitoring data are aligned according to a unified time scale so that all types of data have corresponding values ​​at the same time point. Deformation increment and trend characteristics are extracted from displacement data; stress change characteristics are extracted from stress data; energy release intensity and frequency characteristics are extracted from microseismic data; and pressure change characteristics are extracted from seepage data. Based on spatial location, combined analysis is performed on data from the same cross section and adjacent measuring points to extract spatial difference characteristics; Simultaneously, trend analysis is performed on the time series data to extract the rate of change and fluctuation characteristics.

6. The intelligent prediction method for the stability of surrounding rock in underground coal mine roadways as described in claim 5, characterized in that: The constructed surrounding rock stability risk scoring model includes: The features in the structured feature set are used as inputs for different dimensions, and corresponding weights are set for each dimension; During the calculation process, various features are normalized and then comprehensively weighted to obtain a comprehensive risk score that reflects the current stability of the surrounding rock. Deformation characteristics are used to reflect the degree of deformation of the surrounding rock, stress characteristics are used to reflect the degree of stress concentration, and microseismic characteristics are used to reflect the intensity of potential destructive activity. After the model calculation is completed, the resulting comprehensive risk score is used as the basic assessment result and passed to the inference node of the large language model.

7. The intelligent prediction method for the stability of surrounding rock in underground coal mine roadways as described in claim 6, characterized in that: The output surrounding rock stability level and prediction results include, In the inference node of the workflow, the basic assessment results are fused with the current structured features and retrieved historical cases and normative rules and input into the DeepSeek large language model. The DeepSeek large language model performs semantic correction and interpretation of the risk score and outputs the final stability level and key risk factors. The stability level and trend of change are judged in the early warning judgment node. When the preset level is reached or a continuous deterioration trend occurs, an early warning event is triggered. The generated early warning information includes the risk level, risk source, and handling suggestions, and is sent to the mine dispatch room terminal through the push node in the workflow.

8. A smart prediction system for the stability of surrounding rock in underground coal mine roadways, employing the smart prediction method for the stability of surrounding rock in underground coal mine roadways as described in any one of claims 1 to 7, characterized in that: It includes a data acquisition module, a data upload module, and a prediction module; The data acquisition module is used to collect real-time data streams related to the working conditions of the surrounding rock in the roadway, and to attach a unified timestamp and acquisition location identifier to each data to form a multi-source data set with temporal consistency and spatial correlation, while performing data consistency control and cross-measuring point correlation analysis. The data upload module is used to upload multi-source data sets to the intelligent agent through the data interface, and the Dify workflow orchestration engine performs node-based orchestration and dynamic scheduling execution to build a structured feature set with unified time steps and spatial correlation. The prediction module is used to construct a surrounding rock stability risk scoring model based on a structured feature set, and combines historical cases and normative rules obtained from retrieval as the basis for reasoning. The model is then input into the DeepSeek large language model for reasoning correction, and outputs the surrounding rock stability level and prediction results.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent prediction method for the stability of surrounding rock in underground coal mine roadways as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent prediction method for the stability of surrounding rock in underground coal mine roadways as described in any one of claims 1 to 7.

Images