Mine design knowledge accumulation and sedimentation system and method based on large model technology

Abstract

The application discloses a mine design knowledge accumulation and sedimentation system and method based on a large model technology, relates to the technical field of industrial knowledge automation, and comprises the following steps: when a knowledge optimization process is triggered, collecting historical design cases to construct a space-time knowledge graph, acquiring perforation blasting process parameters, inputting the perforation blasting process parameters and real-time environmental data into an Llama-3 model, executing an environmental response rule processing process, and obtaining an optimized parameter data package; inputting the optimized parameter data package into an LS-DYNA explicit dynamics analysis engine, executing blasting vibration propagation law verification calculation to obtain blasting vibration velocity values and rock mass maximum tensile stress values, and generating a counterexample data set; feeding back the counterexample data set to the Llama-3 model, re-executing the environmental response rule processing process until no counterexample data set is generated, and generating a physical compliance parameter package; and through a closed-loop learning formed by the generation and feedback mechanism of the counterexample data set, the problem of insufficient physical compliance verification is solved.

Description

Technical Field

[0001] This invention relates to the field of industrial knowledge automation technology, and in particular to a system and method for accumulating and preserving mine design knowledge based on large model technology. Background Technology

[0002] With the development of smart mines, digital twin mines, and industrial knowledge automation technologies, mine design knowledge management is gradually evolving from paper archives, manual experience records, and single engineering databases towards multi-source data fusion, engineering knowledge modeling, and intelligent decision support. Current mine designs typically rely on information resources such as geological exploration data, historical blasting cases, drilling blasting process parameters, environmental monitoring data, and engineering safety specifications. They combine these with CAD / BIM modeling, mine design software, knowledge graphs, expert rule bases, machine learning models, and numerical simulation analysis tools to organize, express, and reuse mine operation parameters, environmental conditions, design experience, and safety evaluation results. This supports the generation of mine design schemes, management of process parameters, accumulation of engineering knowledge, and analysis of operational safety.

[0003] While existing methods have made some progress in the digitization and automation of mine design, shortcomings remain. First, existing knowledge accumulation mechanisms lack the ability to model responses to dynamic environments, resulting in lagging knowledge updates and a lack of physical consistency. Second, existing methods generally rely on human experience or simplified assumptions in the knowledge verification stage, failing to fully consider multi-physics coupling effects, leading to insufficient physical compliance of optimization parameters. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for accumulating and preserving mine design knowledge based on large model technology to solve the problems of lagging dynamic knowledge accumulation and insufficient physical compliance verification in mine design.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: Firstly, this invention provides a method for accumulating and preserving mine design knowledge based on large-scale model technology, which includes: Continuously collect and preprocess real-time environmental data; the real-time environmental data includes slope displacement, blasting vibration waveform, rainfall intensity, and rock stress data; Real-time environmental data is substituted into the field effect equation to obtain knowledge confidence. When the knowledge confidence is lower than the confidence threshold, the knowledge optimization process is triggered. When the knowledge optimization process is triggered, historical design cases are collected to construct a spatiotemporal knowledge graph, the perforation blasting process parameters are obtained, the perforation blasting process parameters and real-time environmental data are input into the Llama-3 model, the environmental response rule processing process is executed, and the optimized parameter data package is obtained. The optimized parameter data package is input into the LS-DYNA explicit dynamic analysis engine to perform the verification calculation of the blasting vibration propagation law, obtain the blasting vibration velocity value and the maximum tensile stress value of the rock mass, and generate a set of counterexample data. The set of counterexample data is fed back to the Llama-3 model, and the environmental response rule processing flow is re-executed until no more counterexample data sets are generated, and a physical compliance parameter package is generated. The physical compliance parameter package is encoded according to its structure, and the spatiotemporal knowledge graph is updated using the encoded physical compliance parameters and the corresponding association rule expressions.

[0007] As a preferred embodiment of the mine design knowledge accumulation and sedimentation method based on large model technology described in this invention, the preprocessing refers to upsampling the real-time environmental data, filtering the upsampled real-time environmental data, detecting and deleting abnormal data in the filtered real-time environmental data, and filling in the blanks left by deleting abnormal data.

[0008] As a preferred embodiment of the mine design knowledge accumulation and sedimentation method based on large model technology described in this invention, the specific steps for obtaining knowledge confidence by substituting real-time environmental data into the field effect equation are as follows: Collect historical blasting cases, calculate the initial knowledge confidence based on the historical blasting cases, and use the initial knowledge confidence to construct an exponential decay model; Environmental disturbances are measured by blasting tests with controlled variables. The exponential decay model is multiplicatively coupled with the environmental disturbances to form a field-effect equation. The real-time environmental data is then substituted into the field-effect equation to obtain the knowledge confidence level.

[0009] As a preferred embodiment of the method for accumulating and preserving mine design knowledge based on large model technology described in this invention, the confidence threshold is set by collecting historical operation records and statistically analyzing historical accident rates.

[0010] As a preferred embodiment of the method for accumulating and preserving mine design knowledge based on large model technology described in this invention, the construction of the spatiotemporal knowledge graph refers to collecting historical design cases, defining process parameter nodes and environmental condition nodes, and creating applicability relationship edges.

[0011] As a preferred embodiment of the mine design knowledge accumulation and sedimentation method based on large model technology described in this invention, the specific steps for inputting the perforation blasting process parameters and real-time environmental data into the Llama-3 model are as follows: Read the perforation blasting process parameters from the spatiotemporal knowledge graph; Collect historical mining operation data and use the historical mining operation data to train the Llama-3 model; Based on historical mining operation data, strong association rules are generated using association rule mining algorithms, and then transformed into condition thresholds and adjustment instructions to form an environmental response rule base. The environmental response rule base is loaded into the trained Llama-3 model, and the perforation blasting process parameters and real-time environmental data are input into the trained Llama-3 model to execute the environmental response rule processing flow.

[0012] As a preferred embodiment of the mine design knowledge accumulation and sedimentation method based on large model technology described in this invention, the execution of the environmental response rule processing flow refers to comparing the real-time environmental data with the condition threshold in the environmental response rule base. When the real-time environmental data exceeds the condition threshold, the corresponding adjustment instruction is applied to obtain the optimized parameter data package.

[0013] As a preferred embodiment of the mine design knowledge accumulation and sedimentation method based on large model technology described in this invention, the generation of counterexample data set refers to calculating the critical safety threshold and the critical value of dynamic tensile strength of rock mass, and recording the generation of counterexample data set when the blasting vibration velocity value is lower than the critical safety threshold and the maximum tensile stress value of rock mass is lower than the critical value of dynamic tensile strength of rock mass are not simultaneously satisfied.

[0014] As a preferred embodiment of the mine design knowledge accumulation and sedimentation method based on large model technology described in this invention, the specific steps for encoding the physical compliance parameter package according to its structure are as follows: Encapsulate the physical compliance parameter package as a structured triple; Convert the conditional threshold into the logical expression antecedent and the adjustment instruction into the assignment expression consequent. The antecedent of the logical expression and the consequent of the assignment expression are combined into an association rule expression.

[0015] Secondly, this invention provides a mine design knowledge accumulation and sedimentation system based on large-scale model technology, including, The preprocessing module continuously collects and preprocesses real-time environmental data, including slope displacement, blasting vibration waveform, rainfall intensity, and rock stress data. The trigger module substitutes real-time environmental data into the field effect equation to obtain the knowledge confidence level. When the knowledge confidence level is lower than the confidence threshold, the knowledge optimization process is triggered. When the knowledge optimization process is triggered, the execution module collects historical design cases to build a spatiotemporal knowledge graph, obtains the perforation blasting process parameters, inputs the perforation blasting process parameters and real-time environmental data into the Llama-3 model, executes the environmental response rule processing process, and obtains the optimization parameter data package. The generation module inputs the optimized parameter data package into the LS-DYNA explicit dynamic analysis engine, performs verification calculations on the propagation law of blasting vibration, obtains the blasting vibration velocity value and the maximum tensile stress value of the rock mass, and generates a set of counterexample data. The set of counterexample data is fed back to the Llama-3 model, and the environmental response rule processing flow is re-executed until no more counterexample data sets are generated, and a physical compliance parameter package is generated. The writing module encodes the physical compliance parameter package according to its structure, and uses the encoded physical compliance parameters and corresponding association rule expressions to update the spatiotemporal knowledge graph.

[0016] The beneficial effects of this invention are as follows: By constructing a spatiotemporal knowledge graph, historical cases are dynamically associated with real-time data to form a scalable knowledge network, solving the problem of lagging knowledge updates; by generating and feeding back counterexample data sets, a closed-loop learning mechanism is formed, realizing continuous accumulation and iteration of knowledge, reducing reliance on human experience, improving the safety and scientific nature of the design, and solving the problem of insufficient physical compliance verification; by calculating knowledge confidence through field effect equations, the reliability of the current knowledge base is dynamically evaluated, enabling real-time response to environmental changes and avoiding misjudgments of complex working conditions by existing static evaluation methods. 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 A flowchart for methods of accumulating and refining mine design knowledge based on large model technology.

[0019] Figure 2 This is a schematic diagram of a mine design knowledge accumulation and sedimentation system based on large model technology.

[0020] Figure 3 This is a flowchart for preprocessing.

[0021] Figure 4 A flowchart for constructing the field effect equation and triggering the knowledge optimization process. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for accumulating and preserving mine design knowledge based on large model technology, including the following steps: S1: Continuously collect and preprocess real-time environmental data; real-time environmental data includes slope displacement, blasting vibration waveform, rainfall intensity, and rock stress data.

[0026] The specific steps are as follows: S1.1: In the blasting area of ​​the open-pit mine, a slope displacement monitoring instrument is used to record the three-dimensional coordinate changes of the open-pit mine slope rock mass, a blasting vibration meter is used to record the vibration waveforms of the open-pit mine slope rock mass along the XYZ axes, a meteorological monitoring station is used to record rainfall, and a rock mass stress sensor is used to collect the internal strain values ​​of the rock mass.

[0027] The changes in three-dimensional coordinates, vibration waveforms, rainfall, and internal strain values ​​of the rock mass are stored in the edge computing nodes to form real-time environmental data.

[0028] S1.2: Using the acquisition time axis of the blasting vibration waveform as a reference, the real-time environmental data is upsampled to a fixed frequency using piecewise cubic spline interpolation to maintain the linear characteristics of the slope displacement plateau area and enhance the curvature continuity of the abrupt change area. Piecewise linear interpolation is also used to upsample the rainfall intensity and rock stress data to a fixed frequency.

[0029] A fourth-order Butterworth low-pass filter was used to filter the upsampled slope displacement, rainfall intensity, and rock stress data to identify and filter out interpolation noise components with frequencies higher than a fixed frequency. The cutoff frequency was set to twice the main frequency of the corresponding upsampled real-time environmental data.

[0030] It should also be noted that in this embodiment, the fixed frequency is 1000Hz because 1000Hz can serve as a unified time axis frequency to align low-frequency slope displacement, rainfall intensity, and rock stress data to the blasting vibration waveform time axis. A frequency greater than 1000Hz would increase the number of interpolation points and computational load for slope displacement, rainfall intensity, and rock stress data, while a frequency less than 1000Hz would reduce the alignment accuracy between slope displacement, rainfall intensity, and rock stress data and the blasting vibration waveform time axis, easily causing the loss of abrupt response characteristics.

[0031] S1.3: In the filtered real-time environmental data, detect and delete zero-value segments in the upper slope displacement that last for a fixed time, saturated waveforms in the blasting vibration waveforms whose amplitude reaches the upper limit of the range, negative records in the rainfall intensity, and data points in the rock mass stress data whose jump variables reach the upper limit of the range. Use linear interpolation to fill in the deleted segments and generate corrected real-time environmental data.

[0032] Standardize the units of the corrected real-time environmental data: unify the unit of slope displacement to millimeters, the unit of blasting vibration waveform to centimeters per second, the unit of rainfall intensity to millimeters per hour, and the unit of rock mass stress data to megapascals.

[0033] It should also be noted that preprocessing improves the quality and reliability of real-time environmental data, providing an accurate input basis for subsequent knowledge confidence calculations and process optimization.

[0034] S2: Substitute real-time environmental data into the field effect equation to obtain knowledge confidence. When the knowledge confidence is lower than the confidence threshold, the knowledge optimization process is triggered.

[0035] The specific steps are as follows: S2.1: The effectiveness of mine design knowledge is directly related to production safety and efficiency. Existing methods suffer from the drawbacks of static knowledge failure, lack of experience quantification, and neglect of time decay, which leads to a decrease in the reliability of mine design knowledge under dynamic environmental changes. This invention solves the above problems by constructing a field effect equation to achieve dynamic quantitative evaluation of knowledge confidence.

[0036] The field effect equation is constructed based on the exponential decay model, which conforms to the law of accelerated decay of knowledge value over time in cognitive science. Environmental disturbances are measured by explosion experiments with controlled variables. The environmental disturbances are coupled with the exponential decay model by multiplication. Because sudden environmental changes can cause knowledge to become invalid instantly, multiplication coupling satisfies the knowledge confidence level of zero when the environment deteriorates to an extreme degree. In an ideal environment, the field effect equation degenerates into the exponential decay model.

[0037] Specifically, the expression for the field-effect equation is as follows: ; in, The current knowledge confidence level, with a value range of [0,1]. Let be the initial knowledge confidence level, representing the inherent reliability of knowledge under ideal conditions, with a value range of [0,1]. The time decay coefficient represents the natural rate of decay of knowledge over time. The time interval represents the time from when the knowledge was entered to the present. For environmental disturbances, the value range is [0,1]. This is the corrected real-time environmental data used to calculate environmental disturbances.

[0038] Furthermore, the initial knowledge confidence level is the ratio of the number of successful historical blasting cases to the total number of historical blasting cases. The time decay coefficient is obtained by fitting the relationship curve between time and failure rate using the maximum likelihood estimation method. The failure rate of design parameters under different environmental conditions is determined by blasting tests with controlled variables. Based on the principle that the sum of the failure rate of design parameters and the success rate of design parameters is one, the success rate of design parameters can be obtained, and the success rate of design parameters is used as an environmental disturbance.

[0039] S2.2: Substitute the environmental disturbances corresponding to the real-time environmental data into the field effect equation to obtain the current knowledge confidence level. Statistically calculate the historical accident incidence rate corresponding to different knowledge confidence level intervals in the historical operation records. Use kernel density estimation to fit the curve of accident rate changing with knowledge confidence level. Select the knowledge confidence level corresponding to the inflection point of the curve as the confidence level threshold.

[0040] The current knowledge confidence level is compared with the confidence threshold. If the current knowledge confidence level is not lower than the confidence threshold, the current design parameters are maintained and a "high confidence" status identifier is output. If the current knowledge confidence level is lower than the confidence threshold, a "low confidence" status identifier is generated and the knowledge optimization process is triggered.

[0041] It should also be noted that the field effect equation enables dynamic evaluation of knowledge and can automatically trigger optimization processes based on environmental changes, thereby improving the real-time nature and accuracy of decision-making.

[0042] S3: When the knowledge optimization process is triggered, historical design cases are collected to build a spatiotemporal knowledge graph, the perforation blasting process parameters are obtained, the perforation blasting process parameters and real-time environmental data are input into the Llama-3 model, the environmental response rule processing process is executed, and the optimized parameter data package is obtained.

[0043] The specific steps are as follows: S3.1: Collect historical design cases to construct a spatiotemporal knowledge graph: Define the perforation blasting process parameters in the historical design cases as process parameter nodes, and define the specific rainfall intensity conditions and specific rock mass stress conditions in the historical design cases as environmental condition nodes. Create applicability relationship edges between process parameter nodes and environmental condition nodes, and attach time validity attributes and geographic coordinate attributes to each process parameter node and environmental condition node.

[0044] It should also be noted that the applicability relationship edge represents the directional association between the perforation blasting process parameters corresponding to the process parameter node and the conditions corresponding to the environmental condition node, indicating whether they are valid or invalid. The perforation blasting process parameters are valid if they can be used safely and achieve the expected technical effect under the current environmental conditions, and invalid if they pose safety hazards or cannot achieve the expected technical effect under the current environmental conditions. The parameters for perforated blasting include charge density, hole size, and drilling depth. The time validity attribute refers to the start and end timestamps, and the geographic coordinate attribute refers to the location coordinates of the mine.

[0045] S3.2: Read environmental monitoring records, drilling blasting process parameter adjustment records, and adjusted blasting effect evaluation records from historical mining operation data, and convert them into natural language command-response pairs.

[0046] Freeze all the original parameters of the Llama-3 model, add trainable rank decomposition matrices next to the attention mechanism and feedforward network layers, select the AdamW optimizer as the optimizer, and select the cosine learning rate scheduler as the scheduler.

[0047] It should also be noted that the rank decomposition matrix includes two low-rank matrices. The first low-rank matrix is ​​obtained by initializing using Kaiming uniform initialization, and the second low-rank matrix is ​​initially a zero matrix.

[0048] In this embodiment, the Llama-3 model is selected as the Llama-3-8B-Instruct version, which includes a word segmenter, a word embedding layer, a Transformer decoding layer, a normalization layer, and a vocabulary mapping layer connected in sequence. The Transformer decoding layer includes a normalization layer, a self-attention sub-layer, a residual connection layer, a normalization layer, a feedforward network sub-layer, and a residual connection layer connected in sequence.

[0049] The inherent public parameters of the Llama-3-8B-Instruct version are as follows: 32 Transformer decoding layers, 32 attention heads, 8 key-value heads, 128 dimensions per attention head, 14336 dimensions in the feedforward network, 128256 words, and a maximum position length of 8192.

[0050] The natural language instruction-response pairs are converted into text sequences. The text sequences are then input into the Llama-3 model. A word segmenter is used to segment the text sequences to obtain a sequence of word numbers. The word embedding vector corresponding to each word number is read from the word embedding matrix of the word embedding layer to form a sequence of word embedding vectors.

[0051] The Transformer decoding layer performs root mean square normalization on each word embedding vector to obtain a normalized word embedding vector. The normalized word embedding vector is then multiplied by the query weight matrix, key weight matrix, and value weight matrix, respectively, to obtain the query vector, key vector, and value vector.

[0052] The query weight matrix, key weight matrix, and value weight matrix are the original parameters of the Transformer decoding layer.

[0053] The query vector and key vector are positionally encoded using rotational positional encoding to obtain positionally encoded query vectors and positionally encoded key vectors. The dot product similarity between the positionally encoded query vector and the positionally encoded key vector is calculated to obtain the attention similarity score.

[0054] A causal mask is applied to the attention score and then Softmax normalization is performed to obtain the attention weights. The attention weights are then used to perform a weighted summation of the value vectors to obtain the head output vector of the attention head.

[0055] All head output vectors are concatenated into a multi-head attention output vector. The product of the multi-head attention output vector and the weight matrix of the self-attention sub-layer is used as the attention projection output vector. The sum of the attention projection output vector and the word embedding vector is used as the attention sub-layer output vector.

[0056] The feedforward sublayer performs root mean square normalization on the output vector of the attention sublayer to obtain the feedforward normalized vector.

[0057] The product of the feedforward normalized vector and the gate weight matrix of the feedforward network sublayer is used as the original gated mapping vector. The product of the feedforward normalized vector and the corresponding first low-rank matrix is ​​used as the gated low-rank intermediate vector. The product of the gated low-rank intermediate vector and the corresponding second low-rank matrix is ​​used as the initial gated low-rank mapping vector. The initial gated low-rank mapping vector is scaled according to the ratio of the alpha value to the rank of the Llama-3 model to obtain the gated low-rank incremental mapping result. The sum of the original gated mapping vector and the gated low-rank incremental mapping result is used as the gated vector.

[0058] The product of the feedforward normalized vector and the upprojection weight matrix of the feedforward network sublayer is used as the original upprojection mapping vector. The product of the feedforward normalized vector and the corresponding first low-rank matrix is ​​used as the upprojection low-rank intermediate vector. The product of the upprojection low-rank intermediate vector and the corresponding second low-rank matrix is ​​used as the initial upprojection low-rank mapping vector. The initial upprojection low-rank mapping vector is scaled according to the ratio of the alpha value to the rank of the Llama-3 model to obtain the upprojection low-rank incremental mapping result. The sum of the original upprojection mapping vector and the upprojection low-rank incremental mapping result is used as the upprojection vector.

[0059] Substitute each element of the gated vector into the SiLU activation function to obtain the activated gated vector. Perform a dot product operation between the activated gated vector and the upper projection vector to obtain the gated intermediate vector.

[0060] The product of the gated intermediate vector and the downprojection weight matrix of the feedforward network sublayer is used as the original downprojection mapping vector. The product of the gated intermediate vector and the corresponding first low-rank matrix is ​​used as the downprojection low-rank intermediate vector. The product of the downprojection low-rank intermediate vector and the corresponding second low-rank matrix is ​​used as the initial downprojection low-rank mapping vector. The initial downprojection low-rank mapping vector is scaled according to the ratio of the alpha value to the rank of the Llama-3 model to obtain the downprojection low-rank incremental mapping result. The sum of the original downprojection mapping vector and the downprojection low-rank incremental mapping result is used as the output vector of the feedforward network.

[0061] The sum of the output vector of the feedforward network and the output vector of the attention sublayer is used as the output vector of the current Transformer decoding layer.

[0062] The output vector of the current Transformer decoding layer is input into the subsequent Transformer decoding layers. Normalization, attention calculation, residual addition, feedforward network calculation, and residual addition are repeated until the output vector of the last Transformer decoding layer is obtained.

[0063] The output vector of the last Transformer decoding layer is normalized by root mean square to obtain the final normalized output vector. The product of the final normalized output vector and the word weight matrix of the word mapping layer is used as the word prediction score for each word position. Softmax normalization is performed on the word prediction score to obtain the probability distribution of predicting the next word for each word.

[0064] Using the text sequence as the target output sequence, calculate the cross-entropy loss value between the probability distribution of predicting the next word for each word and the target output sequence. Calculate the partial derivatives of the cross-entropy loss value with respect to the first low-rank matrix and the second low-rank matrix, respectively, as the gradients of the first and second low-rank matrices. Use the AdamW optimizer to update the first and second low-rank matrices according to the gradients of the first and second low-rank matrices, respectively.

[0065] At the end of each training cycle, calculate the cross-entropy loss value and save the checkpoint of the Llama-3 model with the minimum cross-entropy loss value.

[0066] The first and second low-rank matrices obtained from training are combined with the weights of the Llama-3 model to generate the trained Llama-3 model.

[0067] S3.3: Extract environmental monitoring records and corresponding drilling and blasting process parameter adjustment records from historical mining operation data. Use association rule mining algorithms to find frequent multisets and strong association rules between environmental conditions and parameter adjustment amounts, and convert them into condition thresholds and adjustment instructions to form an environmental response rule base. Load the environmental response rule base into the trained Llama-3 model.

[0068] Specifically, environmental monitoring records and corresponding drilling and blasting process parameter adjustment records are aligned by timestamp and merged into a transaction dataset. Environmental monitoring records are used as condition items, drilling and blasting process parameter adjustment records are used as result items, and the relationship between condition items and corresponding result items is used as association rules.

[0069] The Apriori algorithm is used to process transaction datasets: a frequent set of single-item condition items or result items that appear alone is generated, a frequent set of multiple-item condition items and result items that appear simultaneously is generated using combination expansion, and the frequency of occurrence of four combinations of condition items and result items in the transaction dataset is counted and a contingency table is constructed.

[0070] The frequencies of occurrence for the four combinations are as follows: the frequency of both the condition and the result appearing simultaneously, the frequency of the condition appearing but the result not appearing, the frequency of the result appearing but the condition not appearing, and the frequency of neither the condition nor the result appearing.

[0071] The ratio of the co-occurrence frequency of the condition and the result to the occurrence frequency of the condition is used as the confidence level of the association rule. A chi-square test is performed on the contingency table to obtain the p-value used to determine whether the association rule is statistically significant. The ratio of the confidence level of the association rule to the independent occurrence frequency of the result is used as the lift. The lift threshold is set based on the historical lift using the percentile method. For example, the historical lift at the 75th percentile is used as the lift threshold. Association rules with p-values ​​below a fixed value and lift not less than the lift threshold are considered strong association rules.

[0072] Strongly correlated rules are extracted from frequent multisets. The condition terms of the strongly correlated rules are transformed into condition thresholds, and the result terms are transformed into adjustment instructions. For example, "reducing the charge density by 15%" is transformed into "charge density × 0.85", forming an environmental response rule base.

[0073] S3.4: Obtain the drilling and blasting process parameters that are valid for the current time point and applicable to standard environmental conditions from the spatiotemporal knowledge graph. Standard environmental conditions refer to a stable operating state where the rainfall intensity is below a specific value and the rock mass stress is below a specific value.

[0074] The perforation blasting process parameters and real-time environmental data are encapsulated in JSON format and input into the trained Llama-3 model. The environmental response rule processing flow is executed: the trained Llama-3 model calculates the probability distribution of the current word predicting the next word in the JSON format text according to step S3.2, takes the word with the highest probability as the candidate output word, and merges the continuously generated candidate output words into the candidate output text.

[0075] The candidate output text includes the comparison results of real-time environmental data and conditional thresholds, the activation status of association rules, and the optimized parameter values. The optimized parameter values ​​are encapsulated into an optimized parameter data package in JSON format.

[0076] It should also be noted that the Llama-3 model combines historical mining operation data and real-time environmental data to generate optimized parameter data packages, thereby improving the adaptability and security of the knowledge.

[0077] S4: Input the optimized parameter data package into the LS-DYNA explicit dynamic analysis engine, perform the blasting vibration propagation law verification calculation to obtain the blasting vibration velocity value and the maximum tensile stress value of the rock mass, and generate a set of counterexample data. Feed the set of counterexample data back to the Llama-3 model, re-execute the environmental response until no more counterexample data sets are generated, and generate a physical compliance parameter package.

[0078] The specific steps are as follows: S4.1: Convert the optimization parameter data package into K file format. The K file includes the definition of material properties, geometric model parameters, loading curve settings, and boundary condition constraints. Material properties include rock mass density, elastic modulus, and Poisson's ratio. Geometric model parameters refer to the three-dimensional blasting hole array generated based on the charge density value and hole size value in the optimization parameter data package. Loading curve settings refer to defining the explosive detonation pressure time history curve based on the charge density value. The infinite rock mass domain simulated by the non-reflective boundary is used as the boundary condition constraint.

[0079] Specifically, geological surveys are conducted in the mine to collect rock mass density, elastic modulus, and Poisson's ratio to define material properties. The center point coordinates of each blast hole in three-dimensional space are calculated using the mesh size values ​​to form a regular mesh distribution. Hexahedral elements are used to mesh each blast hole and the surrounding rock mass to establish a geometric model.

[0080] Based on the charge density value and explosive type, the detonation velocity and polyhedral index are queried. The charge density value, detonation velocity, and polyhedral index are substituted into the detonation pressure formula to calculate the peak detonation pressure of the explosive. The rise time and exponential decay constant of the explosive detonation pressure time history curve are set. The explosive detonation pressure time history curve is defined using LS-DYNA. The detonation pressure formula is existing technology and will not be elaborated further.

[0081] Apply non-reflective boundary conditions to the outer surface of the geometric model, and combine the material property definitions, geometric model parameters, loading curve settings, and boundary condition constraints into a K file in LS-DYNA format.

[0082] S4.2: The LS-DYNA solver is used to perform explicit dynamic calculations on the geometric model parameters and boundary condition constraints. The calculation process is based on the governing equations of the blasting vibration propagation law, solving the rock mass stress wave propagation process, obtaining the blasting vibration velocity field distribution cloud map and the rock mass stress field distribution cloud map, and extracting the peak value of the blasting vibration velocity and the maximum tensile stress value of the rock mass from the blasting vibration velocity field distribution cloud map and the rock mass stress field distribution cloud map.

[0083] Furthermore, the LS-DYNA solver is used to read the K file and start the solution. The solution process employs explicit central difference method for time integration, and the expression is: ; in, For the first The geometric vertex displacement vector of the mesh at each time step. For the first The geometric vertex displacement vector of the mesh at each time step. For the first The geometric vertex displacement vector of the mesh at each time step. For time step, It is the inverse of the mass matrix. For external load vector, For internal force vectors, For time step indexing.

[0084] The element strain tensor and element stress tensor are calculated at each time step, and the expressions are as follows: ; ; in, For the first The element strain tensor at the time step, The strain-displacement matrix is... For the first The element stress tensor at the time step, For the material stiffness tensor, Represents the relationship between the material stiffness tensor and the first... Tensor double dot product of the element strain tensor at the time step.

[0085] The wave propagation process in the rock mass is calculated based on the stress wave propagation governing equation. The equation of motion is given by Newton's second law, and its expression is: ; in, For rock mass density, For vector differential operators, Force per unit volume.

[0086] Considering the nonlinear strength and failure behavior of the rock mass, the Hoek-Brown yield criterion is used as the plastic model of the material to calculate the maximum effective principal stress of the rock mass under complex stress states. The expression is as follows: ; in, The maximum effective principal stress represents the maximum principal stress acting on the failure surface when the rock mass fails. The minimum effective principal stress represents the minimum principal stress acting on the failure surface when the rock mass fails. The uniaxial compressive strength of the rock block represents the compressive strength measured in a uniaxial compression test of an intact rock block. This is the reduction constant in the Hoek-Brown criterion that takes into account rock mass structure and leveling. This is a constant in the Hoek-Brown criterion that reflects the degree of lithification. The exponent in the Hoek-Brown criterion describes the shape of the criterion expression.

[0087] The vibration velocity data is obtained by differentiating the displacement vector of the geometric vertex of the mesh with respect to displacement time. The vibration velocity data of the geometric vertices of all meshes are extracted to generate a blasting vibration velocity field distribution cloud map. The stress tensor of all elements is extracted to generate a rock mass stress field distribution cloud map.

[0088] S4.3: The distance between the blast hole and the monitoring point is used as the propagation distance. The critical safety threshold is calculated based on the propagation distance using the Sadovsky formula: Specific vibration velocity data is used as the safe allowable particle vibration velocity. Multiple blasting tests are conducted in the mine. After collecting vibration velocity data, regression analysis is performed to fit the Sadovsky formula to obtain the site coefficient. The safe allowable particle vibration velocity and the site coefficient are substituted into the Sadovsky formula to obtain the critical safety threshold. The Sadovsky formula is existing technology and will not be elaborated further.

[0089] Based on the static tensile strength and blasting loading strain rate of the rock mass, the critical value of the dynamic tensile strength of the rock mass is calculated using Griffith's strength theory. The expression is as follows: ; in, This is the critical value of the dynamic tensile strength of the rock mass. The static tensile strength of the rock mass. The dynamic reinforcement coefficient of the material was calibrated using a split Hopkinson bar test. For reference strain rate, the strain rate level of static testing is usually taken. The strain rate is the strain rate for explosive loading.

[0090] The peak value of the blasting vibration velocity is compared with the critical safety threshold, and the maximum tensile stress value of the rock mass is compared with the critical value of the dynamic tensile strength of the rock mass. If the peak value of the blasting vibration velocity is lower than the critical safety threshold and the maximum tensile stress value of the rock mass is lower than the critical value of the dynamic tensile strength of the rock mass, then the optimized parameter data package is deemed compliant.

[0091] If the peak value of the blasting vibration velocity is not lower than the critical safety threshold, and the maximum tensile stress value of the rock mass is lower than the critical value of the dynamic tensile strength of the rock mass, then the optimized parameter data package is marked as a vibration violation, and the peak value of the blasting vibration velocity is taken as the violation value.

[0092] If the peak value of the blasting vibration velocity is lower than the critical safety threshold, and the maximum tensile stress value of the rock mass is not lower than the critical value of the dynamic tensile strength of the rock mass, then the optimized parameter data package is marked as a stress violation, and the maximum tensile stress value of the rock mass is taken as the violation value.

[0093] If the peak value of the blasting vibration velocity is not lower than the critical safety threshold and the maximum tensile stress value of the rock mass is not lower than the critical value of the dynamic tensile strength of the rock mass, then the optimized parameter data package is marked as a double violation, and the peak value of the blasting vibration velocity and the maximum tensile stress value of the rock mass are taken as the violation values.

[0094] The names of the non-compliant parameters, their values, safety thresholds, and compliance recommendations are integrated into a set of counterexample data. The compliance recommendation is a fixed ratio of the critical safety threshold to the critical value of the dynamic tensile strength of the rock mass.

[0095] S4.4: Feed back the set of negative examples to the trained Llama-3 model through the API interface, and adjust the adjustment instructions in the environment response rule base according to the compliance recommendation values ​​in the set of negative examples.

[0096] Re-execute the environmental response rule processing flow to generate a new optimized parameter data package. Repeat steps S4.1 to S4.3 until the optimized parameter data package is compliant. Integrate the compliant optimized parameter data package, the peak value of blasting vibration velocity, and the maximum tensile stress value of rock mass into a physical compliant parameter package.

[0097] It should also be noted that the LS-DYNA explicit dynamics analysis engine validates and optimizes the parameter data package, generates a set of counterexample data and feeds it back to the trained Llama-3 model, forming a closed-loop learning process to ensure physical compliance and iterative optimization.

[0098] S5: Encode the physical compliance parameter package according to its structure, and update the spatiotemporal knowledge graph using the encoded physical compliance parameters and the corresponding association rule expressions.

[0099] The specific steps are as follows: S5.1: Encapsulate the physical compliance parameter package into a structured triple: Treat the type of the perforation blasting process parameter as an entity, combine the value of the perforation blasting process parameter with the corresponding unit into a key-value pair and treat it as an attribute, treat the condition threshold of step S3.3, the time validity attribute of step S3.1 and the geographic coordinate attribute as spatiotemporal constraints, and connect the entity, attribute and spatiotemporal constraints to form a triple.

[0100] S5.2: Convert the condition threshold into a logical expression antecedent, convert the adjustment instruction into an assignment expression consequent, and combine the logical expression antecedent and the assignment expression consequent into an association rule expression.

[0101] The perforation blasting process parameters in the physical compliance parameter package are transformed into process parameter nodes in the spatiotemporal knowledge graph, and the condition thresholds are transformed into environmental condition nodes in the spatiotemporal knowledge graph. Applicability relationship edges are created between process parameter nodes and environmental condition nodes. Time validity attributes and geographic coordinate attributes are added to each process parameter node and environmental condition node. The peak value of blasting vibration velocity and the maximum tensile stress value of rock mass are added to the corresponding applicability relationship edges.

[0102] It should also be noted that the encoding operation enables the structured storage of the physical compliance parameter package, and after being written into the spatiotemporal knowledge graph, it enables the continuous accumulation and traceability of knowledge.

[0103] This embodiment also provides a mine design knowledge accumulation and sedimentation system based on large model technology, including: The preprocessing module continuously collects and preprocesses real-time environmental data, including slope displacement, blasting vibration waveform, rainfall intensity, and rock stress data. The trigger module substitutes real-time environmental data into the field effect equation to obtain the knowledge confidence level. When the knowledge confidence level is lower than the confidence threshold, the knowledge optimization process is triggered. When the knowledge optimization process is triggered, the execution module collects historical design cases to build a spatiotemporal knowledge graph, obtains the perforation blasting process parameters, inputs the perforation blasting process parameters and real-time environmental data into the Llama-3 model, executes the environmental response rule processing process, and obtains the optimization parameter data package. The generation module inputs the optimized parameter data package into the LS-DYNA explicit dynamic analysis engine, performs verification calculations on the propagation law of blasting vibration, obtains the blasting vibration velocity value and the maximum tensile stress value of the rock mass, and generates a set of counterexample data. The set of counterexample data is fed back to the Llama-3 model, and the environmental response rule processing flow is re-executed until no more counterexample data sets are generated, and a physical compliance parameter package is generated. The writing module encodes the physical compliance parameter package according to its structure, and uses the encoded physical compliance parameters and corresponding association rule expressions to update the spatiotemporal knowledge graph.

[0104] This embodiment also provides a computer device applicable to the method of accumulating and preserving mine design knowledge based on large model technology, 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 method of accumulating and preserving mine design knowledge based on large model technology proposed in the above embodiment.

[0105] 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.

[0106] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the method for accumulating and preserving mine design knowledge based on large model technology 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.

[0107] In summary, this invention addresses the problem of lagging knowledge updates by: constructing a spatiotemporal knowledge graph that dynamically links historical cases with real-time data to form a scalable knowledge network; establishing a closed-loop learning mechanism through the generation and feedback of counterexample data sets, enabling continuous knowledge accumulation and iteration, reducing reliance on human experience, improving the safety and scientific rigor of designs, and resolving insufficient physical compliance verification; and calculating knowledge confidence through field-effect equations to dynamically assess the reliability of the current knowledge base, enabling real-time responses to environmental changes and avoiding misjudgments of complex working conditions by existing static evaluation methods.

[0108] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for accumulating and preserving mine design knowledge based on large-scale model technology, characterized by: include, Continuously collect and preprocess real-time environmental data; The real-time environmental data includes slope displacement, blasting vibration waveform, rainfall intensity, and rock mass stress data; Real-time environmental data is substituted into the field effect equation to obtain knowledge confidence. When the knowledge confidence is lower than the confidence threshold, the knowledge optimization process is triggered. When the knowledge optimization process is triggered, historical design cases are collected to construct a spatiotemporal knowledge graph, the perforation blasting process parameters are obtained, the perforation blasting process parameters and real-time environmental data are input into the Llama-3 model, the environmental response rule processing process is executed, and the optimized parameter data package is obtained. The optimized parameter data package is input into the LS-DYNA explicit dynamic analysis engine to perform the verification calculation of the blasting vibration propagation law, obtain the blasting vibration velocity value and the maximum tensile stress value of the rock mass, and generate a set of counterexample data. The set of counterexample data is fed back to the Llama-3 model, and the environmental response rule processing flow is re-executed until no more counterexample data sets are generated, and a physical compliance parameter package is generated. The physical compliance parameter package is encoded according to its structure, and the spatiotemporal knowledge graph is updated using the encoded physical compliance parameters and the corresponding association rule expressions.

2. The method for accumulating and preserving mine design knowledge based on large-scale model technology as described in claim 1, characterized in that: The preprocessing refers to upsampling the real-time environmental data, filtering the upsampled real-time environmental data, detecting and deleting abnormal data in the filtered real-time environmental data, and filling in the blanks left by deleting abnormal data.

3. The method for accumulating and preserving mine design knowledge based on large-scale model technology as described in claim 1, characterized in that: The specific steps for substituting real-time environmental data into the field-effect equation to obtain knowledge confidence are as follows: Collect historical blasting cases, calculate the initial knowledge confidence based on the historical blasting cases, and use the initial knowledge confidence to construct an exponential decay model; Environmental disturbances are measured by blasting tests with controlled variables. The exponential decay model is multiplicatively coupled with the environmental disturbances to form a field-effect equation. The real-time environmental data is then substituted into the field-effect equation to obtain the knowledge confidence level.

4. The method for accumulating and preserving mine design knowledge based on large-scale model technology as described in claim 1, characterized in that: The confidence threshold is set by collecting historical operation records and statistically analyzing historical accident rates.

5. The method for accumulating and preserving mine design knowledge based on large-scale model technology as described in claim 1, characterized in that: The construction of the spatiotemporal knowledge graph refers to collecting historical design cases, defining process parameter nodes and environmental condition nodes, and creating applicability relationship edges.

6. The method for accumulating and preserving mine design knowledge based on large-scale model technology as described in claim 1, characterized in that: The specific steps for inputting the perforation blasting process parameters and real-time environmental data into the Llama-3 model are as follows. Read the perforation blasting process parameters from the spatiotemporal knowledge graph; Collect historical mining operation data and use the historical mining operation data to train the Llama-3 model; Based on historical mining operation data, strong association rules are generated using association rule mining algorithms, and then transformed into condition thresholds and adjustment instructions to form an environmental response rule base. The environmental response rule base is loaded into the trained Llama-3 model, and the perforation blasting process parameters and real-time environmental data are input into the trained Llama-3 model to execute the environmental response rule processing flow.

7. The method for accumulating and preserving mine design knowledge based on large-scale model technology as described in claim 6, characterized in that: The execution environment response rule processing flow refers to comparing real-time environment data with the condition thresholds in the environment response rule base. When the real-time environment data exceeds the condition thresholds, the corresponding adjustment instructions are applied to obtain an optimized parameter data package.

8. The method for accumulating and preserving mine design knowledge based on large-scale model technology as described in claim 1, characterized in that: The generation of counterexample data set refers to calculating the critical safety threshold and the critical value of the dynamic tensile strength of the rock mass. When the blasting vibration velocity value is lower than the critical safety threshold and the maximum tensile stress value of the rock mass is lower than the critical value of the dynamic tensile strength of the rock mass, the counterexample data set is recorded.

9. The method for accumulating and preserving mine design knowledge based on large-scale model technology as described in claim 1, characterized in that: The specific steps for encoding the physical compliance parameter package according to its structure are as follows. Encapsulate the physical compliance parameter package as a structured triple; Convert the conditional threshold into the logical expression antecedent and the adjustment instruction into the assignment expression consequent. The antecedent of the logical expression and the consequent of the assignment expression are combined into an association rule expression.

10. A mine design knowledge accumulation and sedimentation system based on large-scale model technology, and a mine design knowledge accumulation and sedimentation method based on large-scale model technology as described in any one of claims 1 to 9, characterized in that: include, The preprocessing module continuously collects real-time environmental data and performs preprocessing. The real-time environmental data includes slope displacement, blasting vibration waveform, rainfall intensity, and rock mass stress data; The trigger module substitutes real-time environmental data into the field effect equation to obtain the knowledge confidence level. When the knowledge confidence level is lower than the confidence threshold, the knowledge optimization process is triggered. When the knowledge optimization process is triggered, the execution module collects historical design cases to build a spatiotemporal knowledge graph, obtains the perforation blasting process parameters, inputs the perforation blasting process parameters and real-time environmental data into the Llama-3 model, executes the environmental response rule processing process, and obtains the optimization parameter data package. The generation module inputs the optimized parameter data package into the LS-DYNA explicit dynamic analysis engine, performs verification calculations on the propagation law of blasting vibration, obtains the blasting vibration velocity value and the maximum tensile stress value of the rock mass, and generates a set of counterexample data. The set of counterexample data is fed back to the Llama-3 model, and the environmental response rule processing flow is re-executed until no more counterexample data sets are generated, and a physical compliance parameter package is generated. The writing module encodes the physical compliance parameter package according to its structure, and uses the encoded physical compliance parameters and corresponding association rule expressions to update the spatiotemporal knowledge graph.

Images