A deep learning-based multi-source heterogeneous information fusion method for different types of rock burst incubation processes

Abstract

The present application relates to the technical field of underground engineering rock burst monitoring and deep learning information fusion, and discloses a multi-source heterogeneous information fusion method for different types of rock burst incubation process based on deep learning, which solves the problem that in the rock burst incubation process of underground engineering, multi-source monitoring information is difficult to be fused in the model according to the relationship between energy accumulation and energy release stage, resulting in the inability to accurately match the rock burst type and incubation stage. The present application first constructs a precursor variable sequence by time series alignment of multi-source data, and calculates the energy accumulation and release stage state; with the help of the unified prototype differential attention layer of the prototype differential attention neural network, the positive and negative affinity channels are reconstructed to distinguish the matched and unmatched precursor variables; through prototype attribution update and stage-adaptive fusion, the incubation state fusion representation is obtained, and finally the matched rock burst type and incubation stage joint result are determined and output simultaneously.

Description

Technical Field

[0001] This invention relates to the field of underground engineering rockburst monitoring and deep learning information fusion technology, specifically to a method for fusing multi-source heterogeneous information on different types of rockburst incubation processes based on deep learning. Background Technology

[0002] During the excavation of deep-buried tunnels, hydroelectric caverns, underground mines, and other underground engineering projects, the surrounding rock is subjected to high ground stress and strong disturbance. The elastic energy within the hard, brittle rock mass gradually accumulates with excavation unloading, structural plane expansion, and local fracturing, and is rapidly released at a certain stage, leading to rockburst disasters. The rockburst incubation process is usually accompanied by multiple types of precursory information: microseismic data reflects micro-fracture activity in the rock mass, acoustic emission data reflects crack propagation and energy release, stress data reflects the conditions for elastic energy accumulation, and displacement data reflects the deformation response of the surrounding rock. Because data from different sources have differences in sampling frequency, physical meaning, and scale of change, it is necessary to uniformly process multi-source heterogeneous monitoring information at the engineering site to provide a data foundation for identifying rockburst types and determining the incubation stage.

[0003] Chinese patent CN117972852A discloses a dynamic prediction method for rockburst based on multi-source heterogeneous data and machine learning. This method first conducts numerical simulations of tunnel excavation, using in-situ stress, surrounding rock physical and mechanical parameters, and joints and fissures as inputs, and rockburst energy dissipation rate as output to form numerical samples, which are then trained using a machine learning model. During tunnel excavation, in-situ stress is inverted based on real-time monitoring of microseismic data and acoustic emission data, combined with dynamically updated surrounding rock physical and mechanical parameters and joints and fissures, to predict the energy dissipation rate of the surrounding rock ahead and obtain the rockburst level. Simultaneously, rockburst levels are predicted based on similar engineering data, microseismic data, and acoustic emission data, respectively. Then, initial weights are assigned to multiple rockburst levels and summed to obtain a comprehensive rockburst level.

[0004] While the aforementioned scheme utilizes multi-source heterogeneous data, the multi-source information is primarily weighted and aggregated at the level of rockburst severity or energy dissipation rate. For different types of rockburst incubation processes, there is a clear sequential relationship between elastic energy accumulation, microseismic energy release, acoustic emission energy release, and displacement response. This phase correspondence is not embedded into the model's feature affinity, phase prototype attribution, and fusion pathways. Therefore, it is easy to merge precursor variables corresponding to different phases with those not corresponding to different phases into the same fusion basis, making it difficult to express the inherent correspondence between rockburst types and incubation stages within the same excavation time window. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a multi-source heterogeneous information fusion method for different types of rockburst incubation processes based on deep learning, which solves the problem that multi-source monitoring information is difficult to fuse within the model according to the relationship between energy accumulation and energy release stages during the rockburst incubation process in underground engineering, resulting in the inability to accurately match rockburst type and incubation stage.

[0006] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0007] A method for fusing multi-source heterogeneous information on different types of rockburst incubation processes based on deep learning includes the following steps:

[0008] S1. Acquire microseismic data, acoustic emission data, stress data, and displacement data within the time window of underground engineering excavation to form a sequence of monitoring variables for precursors of rockburst incubation;

[0009] S2. Determine the energy accumulation and release stage state based on the monitoring variable sequence of rockburst precursors. The energy accumulation and release stage state is used to represent the correspondence between elastic energy accumulation and microseismic and acoustic emission energy release.

[0010] S3. Based on the energy accumulation and release stage state, the unified prototype differential attention layer in the prototype differential attention neural network, which is used to calculate the affinity relationship between precursor monitoring variables and the prototype of the rockburst incubation stage, is reconstructed with positive and negative affinity channels to form energy stage-constrained positive and negative affinity channels. This enables the precursor monitoring variables corresponding to the stage to form a positive affinity relationship with the prototype of the rockburst incubation stage, and enables the precursor monitoring variables that do not correspond to the stage to form a negative affinity relationship with the prototype of the rockburst incubation stage.

[0011] S4. Based on the energy stage constraint of positive and negative affinity channels, perform prototype attribution calculation on the monitoring variable sequence of rockburst incubation precursors to form a prototype representation of the rockburst incubation stage;

[0012] S5. Based on the prototype representation of the rockburst incubation stage, multi-source heterogeneous information is fused to obtain the fused representation of the rockburst incubation state;

[0013] S6. Determine the rockburst type and incubation stage based on the fusion representation of rockburst incubation state, and output the fusion results of multi-source heterogeneous information of different types of rockburst incubation processes.

[0014] This scheme first performs time-series alignment and serialization processing on four types of heterogeneous monitoring data: microseismic, acoustic emission, stress, and displacement. This unifies the time and field arrangement logic of multi-source data, eliminating fusion obstacles caused by inconsistencies in the dimensions and sampling frequencies of heterogeneous data. Next, it quantifies and constructs the energy accumulation and release stage states based on the rock mass energy evolution law, transforming the physical time-series mechanism of rockburst incubation into network-computable constraints. A prototype differential attention neural network is used to differentiate and reconstruct the internal affinity channels, distinguishing between matching and mismatched monitoring precursors at the model feature calculation layer. This avoids the problem of traditional methods only weighting and fusing at the final result layer, thus confusing the characteristics of different incubation stages. Based on positive and negative affinity constraints, prototype attribution is updated, ensuring that the prototypes at each stage conform to the actual energy evolution law of the surrounding rock. Then, the fusion pathway is adaptively adjusted according to the response intensity of various monitoring data at different stages, achieving staged and refined multi-source information fusion. Finally, the scheme synchronously outputs the corresponding rockburst type and incubation stage determination results. This scheme integrates rockburst energy evolution priors throughout the entire process of data processing, network feature extraction, multi-source information fusion, and state determination. It effectively avoids interference from the cross-mixing of precursor features at different incubation stages, improves the accuracy and temporal continuity of rockburst type identification and incubation stage division, and is suitable for real-time monitoring and early warning of rockbursts in underground engineering projects with high ground stress, such as deep-buried tunnels and underground mines.

[0015] The beneficial effects of this invention are:

[0016] (1) This invention constructs the energy accumulation and release stage state based on microseismic, acoustic emission, stress and displacement monitoring data, and transforms the rock burst incubation time sequence law of "stress accumulation first and then energy release after rupture" into a quantitative constraint that can be input into the neural network. By reconstructing the positive and negative affinity channels through the unified prototype differential attention layer of the prototype differential attention neural network, the precursor monitoring variables that match and do not match the incubation stage are directly distinguished within the model. This avoids the mutual interference of monitoring features of different stages from the source of feature fusion, and solves the defect that the traditional scheme cannot characterize the inherent correspondence between multi-source data and the rock burst incubation stage.

[0017] (2) This invention performs time alignment, fixed field arrangement and time sequence processing on four types of monitoring data of microseismic, acoustic emission, stress and displacement with different sampling frequencies and different physical dimensions, and constructs a regular and unified sequence of precursor monitoring variables. There is no need to design feature extraction rules for each type of monitoring data, which greatly reduces the difficulty of pre-processing for multi-source heterogeneous data fusion and is suitable for the actual engineering scenario of simultaneous monitoring of multiple devices in underground engineering sites.

[0018] (3) This invention uses positive and negative affinity channels constrained by energy stage to complete the prototype attribution difference calculation. It uses positive affinity to strengthen the contribution of matching variables to the stage prototype and negative affinity to exclude irrelevant variables, and dynamically aggregates and updates the prototypes of each rockburst incubation stage. The prototype representation fits the actual energy evolution process of the surrounding rock. The final output of the incubation stage division time sequence jumps less and the stage boundaries fit the actual evolution law on site, thus improving the accuracy of stage identification.

[0019] (4) Based on the source response intensity of microseismic, acoustic emission, stress and displacement data in each stage prototype, this invention adaptively retains the fusion path of high contribution monitoring data and suppresses the low contribution invalid path, realizes the phased differentiated information fusion, avoids the feature dilution problem caused by the indiscriminate weighting of multi-source data, and gives full play to the early warning advantages of different monitoring methods in different stages of rockburst incubation.

[0020] (5) Based on a unified fusion representation of rockburst incubation state, this invention outputs two types of judgment results in parallel: rockburst type and incubation stage, and establishes a one-to-one joint relationship between the two. There is no need to model and identify rockburst type and evolution stage separately twice. The output results can fully reflect the current evolution stage of the surrounding rock and the rockburst disaster category, and can directly provide a complete decision basis for rockburst classification early warning and support treatment of high ground stress projects such as deep buried tunnels, underground mines, and underground caverns of hydropower. Attached Figure Description

[0021] Figure 1 This is a flowchart of the multi-source heterogeneous information fusion method for different types of rockburst incubation processes based on deep learning in this invention.

[0022] Figure 2 This is a flowchart illustrating the process of constructing a monitoring variable sequence for precursors to rockbursts in this invention.

[0023] Figure 3 This is a flowchart for determining the state of the energy accumulation and release stage in this invention.

[0024] Figure 4 This is a flowchart illustrating the reconstructed energy stage constraint of positive and negative affinity channels in this invention.

[0025] Figure 5 This is a flowchart illustrating the prototype representation of the rockburst incubation stage in this invention.

[0026] Figure 6 This is a flowchart illustrating the process of generating a fusion representation of the rockburst incubation state in this invention.

[0027] Figure 7 This is a flowchart illustrating the joint determination of rockburst type and incubation stage in this invention.

[0028] Figure 8This is a schematic diagram of the overall process of prototype differential fusion under the constraint of the energy accumulation and release stage in this invention.

[0029] Figure 9 This is a schematic diagram of the time alignment of multi-source monitoring variables and the construction of the prototype differential attention neural network input in this invention.

[0030] Figure 10 This is a heatmap comparing the affinity of the prototype before and after applying energy stage constraints in this invention.

[0031] Figure 11 This is a schematic diagram showing the comparison between the predicted rockburst incubation stage and the actual stage markings in the field.

[0032] Figure 12 This is a schematic diagram illustrating the logic for jointly determining rockburst type and incubation stage in this invention. Detailed Implementation

[0033] This invention aims to provide a deep learning-based method for fusing multi-source heterogeneous information on different types of rockburst incubation processes. This addresses the problem that in underground engineering rockburst incubation processes, multi-source monitoring information is difficult to fuse within the model according to the relationship between energy accumulation and energy release stages, leading to an inability to accurately match rockburst types and incubation stages. Its core idea is:

[0034] First, the four types of heterogeneous monitoring data—microseismic, acoustic emission, stress, and displacement—are uniformly aligned in time series to construct a standardized time-series monitoring variable sequence. Then, based on the rock mass energy evolution mechanism, the energy accumulation and release stage states are quantitatively calculated, transforming the physical time series law of rockburst—"stress accumulation first, followed by rupture energy release"—into a network-recognizable quantitative constraint. A prototype differential attention neural network is introduced, and within the network, a unified prototype differential attention layer reconstructs positive and negative affinity dual channels based on stage states, distinguishing between matching and mismatched precursor monitoring variables, and eliminating irrelevant feature interference from the feature fusion source. Based on positive and negative affinity constraints, prototype attribution calculations are completed, and the prototypes for each rockburst incubation stage are dynamically updated. Then, based on the response strength of various monitoring data at different stages, the fusion pathway is adaptively adjusted to complete the staged, refined multi-source information fusion. Finally, based on the unified fusion features, the joint determination of rockburst type and incubation stage is completed simultaneously, establishing a one-to-one matching relationship between the two, achieving intelligent identification of the entire rockburst process that conforms to the actual evolution law of the rock mass.

[0035] To facilitate a better understanding of the technical solutions of this invention, some of the terms involved in this invention are explained below:

[0036] Time position: refers to the chronological position of each sampling time within the underground engineering excavation time window.

[0037] Monitoring variables at the sampling time: refers to the combination of monitoring variables extracted from microseismic data, acoustic emission data, stress data, and displacement data within the same sampling time.

[0038] Fixed source order: This refers to arranging the monitoring variables corresponding to microseismic data, acoustic emission data, stress data, and displacement data in the same source order at each sampling time.

[0039] Aligned monitoring variables: refers to the group of monitoring variables at the sampling time after time alignment and fixed arrangement according to the source relationship.

[0040] Time-source variable group: refers to the combination of variables formed by matching microseismic data, acoustic emission data, stress data and displacement data in the monitoring variable sequence of rockburst precursors at the same sampling time.

[0041] Stress growth direction, stress growth amplitude, and stress growth duration: These refer to the direction and amplitude of stress data change between adjacent sampling times, as well as the duration for which the growth conditions are continuously met.

[0042] Elastic energy accumulation benchmark: refers to the criteria used to characterize the elastic energy accumulation process, which are determined based on the direction of stress growth, the magnitude of stress growth, and the duration of stress growth.

[0043] Candidate sequence of stage states: refers to the candidate sequence of stage states arranged according to the sampling time and containing the elastic energy accumulation benchmark.

[0044] Microseismic energy release change: refers to the change in microseismic energy release calculated based on the changes in microseismic data between adjacent sampling times.

[0045] Microseismic constraint stage state candidate sequence: refers to the stage state candidate sequence that is updated by matching the changes in microseismic energy release with the elastic energy accumulation benchmark in the stage state candidate sequence according to the sampling time.

[0046] Acoustic emission energy release change: refers to the change in acoustic emission energy release calculated based on the changes in acoustic emission data between adjacent sampling times.

[0047] Microseismic and acoustic emission constrained stage state candidate sequence: refers to the stage state candidate sequence that is updated by matching the changes in acoustic emission energy release with the elastic energy accumulation benchmark in the microseismic constrained stage state candidate sequence according to the sampling time.

[0048] Consistency determination of sampling time: refers to determining the correspondence between displacement response changes and candidate sequences of microseismic and acoustic emission constraint stages at the same sampling time.

[0049] Displacement constraint stage state candidate sequence: refers to the stage state candidate sequence formed after updating the microseismic and acoustic emission constraint stage state candidate sequences based on the consistency of sampling time.

[0050] Stage correspondence: refers to the correspondence formed when the elastic energy accumulation benchmark and the changes in microseismic energy release and acoustic emission energy release meet the first preset threshold condition at the same sampling time.

[0051] Stage mismatch: refers to the mismatch formed when the elastic energy accumulation benchmark and the changes in microseismic energy release and acoustic emission energy release at the same sampling time do not meet the first preset threshold condition.

[0052] Energy accumulation and release sequence: refers to the sequence formed by the corresponding and non-corresponding relationships of stages arranged according to the sampling time.

[0053] Energy accumulation and release stage constraint: refers to constraining the affinity channel generation position inside the unified prototype differential attention layer according to the energy accumulation and release stage state, so that the precursor monitoring variables corresponding to the stage form a positive affinity relationship, and the precursor monitoring variables that do not correspond to the stage form a negative affinity relationship.

[0054] Prototype Differential Attention Neural Network: This refers to a deep learning network for multi-source time-series monitoring data of rockbursts in underground engineering. The network has multiple prototypes of rockburst incubation stages built in. It can receive the energy accumulation and release stage status calculated from multi-source monitoring data. It reconstructs positive and negative affinity channels through a unified prototype differential attention layer, distinguishes between precursor monitoring variables that match and do not match the incubation stage, and outputs the joint judgment result of rockburst type and incubation stage in parallel after prototype attribution update and phased adaptive fusion.

[0055] Unified Prototype Differential Attention Layer: This refers to the processing layer in the prototype differential attention neural network used to generate the affinity relationship between precursor monitoring variables and prototypes in the rockburst incubation stage, and to reconstruct the positive and negative affinity channels of the affinity relationship.

[0056] Affinity channel generation location: refers to the location where the affinity relationship between the precursor monitoring variables generated within the unified prototype differential attention layer and the prototype of the rockburst incubation stage is calculated.

[0057] Precursor monitoring variables: These refer to monitoring variables formed from microseismic data, acoustic emission data, stress data, and displacement data at the sampling time and included in the sequence of precursor monitoring variables for rockburst incubation.

[0058] Rockburst incubation stage prototype: refers to the prototype object in the unified prototype differential attention layer used to carry the attribution relationship of different incubation stages.

[0059] Initial affinity: refers to the affinity generated by the unified prototype differential attention layer based on the data proximity between the precursor monitoring variables and the prototype in the rockburst incubation stage.

[0060] Affinity relation to be reconstructed: refers to the initial affinity relation selected for reconstructing positive and negative affinity channels after receiving the energy accumulation and release phase state at the affinity channel generation location.

[0061] Initial affinity of stage constraints: refers to the affinity formed after mapping the correspondence in the energy accumulation and release stage state to the affinity to be reconstructed.

[0062] Stage correspondence value: refers to the value calculated based on the initial affinity relationship of the stage constraint, used to indicate whether the precursor monitoring variables correspond to the prototype of the rockburst incubation stage in terms of elastic energy accumulation and microseismic and acoustic emission energy release.

[0063] Affinity polarity label set: refers to the set of labels consisting of stage-corresponding labels and stage-non-corresponding labels, used to determine the affinity direction between precursor monitoring variables and rockburst incubation stage prototypes.

[0064] Positive affinity: refers to the direction of affinity formed between the precursor monitoring variables corresponding to the stage and the prototype of the rockburst incubation stage.

[0065] Negative affinity: refers to the direction of affinity formed between precursor monitoring variables that do not correspond to the stage and the prototype of the rockburst incubation stage.

[0066] Energy phase constraint affinity: refers to the affinity relationship formed after the initial affinity relationship is constrained by positive and negative affinity relationships during the update phase.

[0067] Energy stage constrained positive and negative affinity channels: refers to the affinity channels formed after separating positive and negative channels according to the energy stage constrained affinity relationship.

[0068] Affinity read position in prototype attribution calculation: refers to the position of positive and negative affinity relationships in the positive and negative affinity channels constrained during the read energy stage of prototype attribution calculation.

[0069] Pairing relationship to be assigned: refers to the relationship formed after pairing the precursor monitoring variables in the rockburst incubation precursor monitoring variable sequence with the prototype of the rockburst incubation stage according to the sampling time.

[0070] Energy-constrained pairing relationship: refers to the pairing relationship formed after the reading results from the positive and negative affinity channels of the energy stage are written into the assignment determination criteria of the pairing relationship.

[0071] Positive attribution response: refers to the attribution response of precursor monitoring variables with positive affinity relative to the prototype of the rockburst incubation stage.

[0072] Negative repulsion response: refers to the repulsion response calculated relative to the prototype of the rockburst incubation stage by a precursor monitoring variable with a negative affinity.

[0073] Stage-constrained attribution response: refers to the attribution response obtained by differential calculation of the positive attribution response and negative rejection response of the same precursor monitoring variable relative to the prototype of the same rockburst incubation stage.

[0074] Stage constraint attribution response set: refers to the set of stage constraint attribution responses of each precursor monitoring variable relative to the prototype of each rockburst incubation stage.

[0075] Stage constraint prototype attribution relationship: refers to the relationship determined by the stage constraint attribution response set, in which precursor monitoring variables are assigned to the corresponding rockburst incubation stage prototypes and excluded prototypes that do not correspond to the rockburst incubation stage.

[0076] Prototype aggregation representation: refers to the representation formed by aggregating the precursor monitoring variables assigned to each rockburst incubation stage prototype according to the stage constraint prototype attribution relationship.

[0077] Rockburst incubation stage prototype representation: refers to the stage prototype representation formed after updating the rockburst incubation stage prototype based on the prototype aggregation representation.

[0078] Stage source: refers to the gestation stage source corresponding to the prototype of the rockburst gestation stage to which the precursor monitoring variable belongs.

[0079] Stage source identification prototype representation: refers to the prototype representation formed after the stage source is identified in the prototype representation of rockburst incubation stage.

[0080] Source response: refers to the response results of microseismic data, acoustic emission data, stress data, and displacement data in the same prototype during the same rockburst incubation stage.

[0081] Fusion criteria: refers to the criteria used to determine how microseismic data, acoustic emission data, stress data, and displacement data participate in the fusion process in the same prototype during the same rockburst incubation stage.

[0082] Source response constraint prototype representation: refers to the prototype representation formed after updating the fusion basis with the source response.

[0083] Intra-stage fusion pathway: refers to the pathway through which microseismic data, acoustic emission data, stress data, and displacement data within the same rockburst incubation stage of the prototype are fused.

[0084] Intra-phase fusion representation: refers to the fusion representation formed after preserving or suppressing intra-phase fusion pathways based on the source response.

[0085] Stage relationships between prototypes of rockburst incubation stages: refers to the relationship between prototypes of different rockburst incubation stages according to the order in which they were formed during the incubation stages.

[0086] Rockburst incubation state fusion representation: refers to the fusion representation obtained by fusing multi-source heterogeneous information from the fusion representations corresponding to the prototypes of each rockburst incubation stage.

[0087] Judgment response set: refers to the set of response results calculated based on the fusion representation of rockburst incubation state, corresponding to rockburst type and incubation stage respectively.

[0088] Maximum response position: refers to the output position in the response set where the response result is the largest.

[0089] Combination matching: refers to matching the rockburst type determination result with the incubation stage determination result.

[0090] Joint judgment result: refers to the judgment result formed by combining the rockburst type judgment result and the incubation stage judgment result.

[0091] Multi-source heterogeneous information fusion results for different types of rockburst incubation processes: refers to the output results including rockburst type determination results, incubation stage determination results, and the correspondence between rockburst type and incubation stage.

[0092] In practical implementation, the implementation process of the deep learning-based multi-source heterogeneous information fusion method for different types of rockburst incubation processes provided by this invention is as follows: Figure 1 This includes the following steps:

[0093] S1. Acquire microseismic data, acoustic emission data, stress data, and displacement data within the time window of underground engineering excavation to form a sequence of monitoring variables for precursors of rockburst incubation;

[0094] The purpose of this step is to unify and integrate four types of heterogeneous monitoring data—microseismic, acoustic emission, stress, and displacement—that differ in sampling frequency and physical dimensions within the excavation time window, complete time matching and standardization processing, generate a time-ordered sequence of monitoring variables for precursors to rockburst incubation, and eliminate the obstacles to subsequent fusion calculations caused by the heterogeneity of multi-source data.

[0095] In one exemplary implementation, the process for constructing the monitoring variable sequence of rockburst precursors in this step is as follows: Figure 2 As shown, it includes the following sub-steps:

[0096] S11. Acquire microseismic data, acoustic emission data, stress data, and displacement data within the underground engineering excavation time window, and determine the time position of the microseismic data, acoustic emission data, stress data, and displacement data within the underground engineering excavation time window;

[0097] S12. Based on the time location, match the microseismic data, acoustic emission data, stress data and displacement data corresponding to the same sampling time to form a monitoring variable group for the sampling time;

[0098] S13. Based on the monitoring variable group at the sampling time, set a fixed arrangement position for the source relationship of microseismic data, acoustic emission data, stress data and displacement data, and arrange the monitoring variable group at the sampling time according to the fixed arrangement position of the source relationship to form an aligned monitoring variable group;

[0099] S14. Connect and align the monitoring variable groups according to the time sequence of the underground engineering excavation time window to form a monitoring variable sequence for precursors of rockburst.

[0100] S2. Determine the energy accumulation and release stage state based on the monitoring variable sequence of rockburst precursors. The energy accumulation and release stage state is used to represent the correspondence between elastic energy accumulation and microseismic and acoustic emission energy release.

[0101] The purpose of this step is to use multi-source time-series monitoring data to quantitatively characterize the temporal matching laws of rock elastic energy accumulation, microseismic / acoustic emission energy release, and displacement deformation response, and to transform the physical mechanism of rockburst incubation—"accumulation before release"—into a quantitative constraint state that can be input into a neural network, providing a priori basis for subsequent internal feature constraints of the network.

[0102] The core idea is as follows: When forming the energy accumulation and release sequence, firstly, the sampling time interval for continuous stress data growth is determined based on the elastic energy accumulation benchmark in the stage state candidate sequence. Then, microseismic energy release changes and acoustic emission energy release changes are matched in adjacent sampling times after the sampling time interval. When the microseismic energy release change or acoustic emission energy release change appears before the elastic energy accumulation benchmark, the corresponding sampling time is marked as a stage non-correspondence relationship. When the microseismic energy release change and acoustic emission energy release change appear after the elastic energy accumulation benchmark, and the displacement response change meets the first preset threshold condition, the corresponding sampling time is marked as a stage correspondence relationship.

[0103] In one exemplary implementation, the process for determining the energy accumulation and release phase state in this step is as follows: Figure 3 As shown, it includes the following sub-steps:

[0104] S21. Receive the monitoring variable sequence of rockburst precursors, and match the microseismic data, acoustic emission data, stress data and displacement data at the same sampling time according to the time sequence of the underground engineering excavation time window to form a time source variable group;

[0105] S22. Calculate the stress growth direction, stress growth amplitude, and stress duration between adjacent sampling times based on the stress data in the time source variable group, and map the stress growth direction, stress growth amplitude, and stress duration to the elastic energy accumulation benchmark to form a stage state candidate sequence;

[0106] S23. Calculate the microseismic energy release change between adjacent sampling times based on the microseismic data in the time source variable group, match the microseismic energy release change with the elastic energy accumulation benchmark in the stage state candidate sequence according to the sampling time, update the stage state candidate sequence according to the matching result, and form a microseismic constraint stage state candidate sequence.

[0107] S24. Calculate the change in acoustic emission energy release between adjacent sampling times based on the acoustic emission data in the time source variable group. Match the change in acoustic emission energy release with the elastic energy accumulation benchmark in the microseismic constraint stage state candidate sequence according to the sampling time. Update the microseismic constraint stage state candidate sequence according to the matching result to form the microseismic and acoustic emission constraint stage state candidate sequences.

[0108] S25. Calculate the displacement response change between adjacent sampling times based on the displacement data in the time source variable group, determine the consistency of the displacement response change with the microseismic and acoustic emission constraint stage state candidate sequences in terms of sampling time, update the microseismic and acoustic emission constraint stage state candidate sequences according to the determination results, and form the displacement constraint stage state candidate sequences.

[0109] S26. The elastic energy accumulation benchmark and the changes in microseismic energy release and acoustic emission energy release at each sampling moment in the candidate sequence of displacement constraint stage states are calibrated to correspond to the changes in elastic energy accumulation benchmark, microseismic energy release, and acoustic emission energy release. The calibration results are compared with the first preset threshold condition. If the first preset threshold condition is met, it is marked as a stage correspondence. If the first preset threshold condition is not met, it is marked as a stage non-correspondence, thus forming an energy accumulation and release correspondence sequence.

[0110] S27. Determine the energy accumulation and release stage state according to the corresponding sequence of energy accumulation and release, so that the energy accumulation and release stage state represents the correspondence between elastic energy accumulation and micro-vibration and acoustic emission energy release.

[0111] S3. Based on the energy accumulation and release stage state, the unified prototype differential attention layer in the prototype differential attention neural network, which is used to calculate the affinity relationship between precursor monitoring variables and the prototype of the rockburst incubation stage, is reconstructed with positive and negative affinity channels to form energy stage-constrained positive and negative affinity channels. This enables the precursor monitoring variables corresponding to the stage to form a positive affinity relationship with the prototype of the rockburst incubation stage, and enables the precursor monitoring variables that do not correspond to the stage to form a negative affinity relationship with the prototype of the rockburst incubation stage.

[0112] The purpose of this step is to introduce the energy accumulation and release phase state into the core layer of the prototype differential attention neural network, reconstruct the positive and negative affinity channels, establish positive affinity for precursor variables that match the gestation phase and negative repulsion affinity for unmatched variables, isolate irrelevant monitoring features from the source of model feature calculation, and avoid mixing of precursor information from different gestation phases.

[0113] Its core principle is as follows: When updating the affinity direction in the initial affinity relationship of the stage constraint according to the affinity polarity label set, the energy accumulation and release stage state is limited to the affinity channel generation position within the unified prototype differential attention layer; for precursor monitoring variables and rockburst incubation stage prototypes whose initial affinity relationship meets the preset positive threshold and are marked as stage non-corresponding labels by the affinity polarity label set, they are updated to negative affinity relationships; for precursor monitoring variables and rockburst incubation stage prototypes marked as stage corresponding labels by the affinity polarity label set, they are updated to positive affinity relationships, and the updated positive and negative affinity relationships are used as the basis for forming positive and negative affinity channels constrained by the energy stage.

[0114] In one exemplary implementation, the prototype differential attention neural network includes: an input mapping layer, a time window encoding layer, a unified prototype differential attention layer, a prototype attribution calculation layer, a multi-source heterogeneous information fusion layer, and an output decision layer.

[0115] The input mapping layer is used to receive the monitoring variable sequence of rockburst precursors. The monitoring variable at a single sampling time in the sequence is a 12-dimensional vector. The network contains 64 fully connected neurons. The activation function is ReLU. The input vector is subjected to linear transformation and nonlinear activation, and the output is a 64-dimensional variable representation.

[0116] The time window encoding layer is used to receive 64-dimensional variable representations, with 128 fully connected neurons and a ReLU activation function built in. It performs temporal feature mapping and encoding on the variable representations to generate 128-dimensional temporal features with uniform dimensions.

[0117] The unified prototype differential attention layer has 6 sets of 128-dimensional rockburst incubation stage prototypes built in. First, it calculates the initial affinity relationship between the temporal features and the prototypes of each stage through a fully connected layer. Then, it receives the four-dimensional energy accumulation and release stage state vector from the external input, generates stage constraint results through affinity reconstruction calculation, compares the corresponding thresholds of the stages to divide affinity polarity labels, and generates two sets of N×6-dimensional channel matrices, positive affinity channel and negative affinity channel, based on the label separation to complete the reconstruction of positive and negative affinity channels.

[0118] The prototype attribution calculation layer reads the positive and negative affinity channel strengths, calculates the positive attribution response and negative rejection response of the prototype corresponding to each stage of the precursor monitoring variable, and performs a difference operation on the temporal features and the similarity of the stage prototype vector to obtain the stage constraint attribution response; based on the rejection threshold, mismatched prototypes are screened out, and based on the attribution threshold, the variable prototype is assigned. The temporal features attributed to the same prototype are weighted and aggregated with the attribution response as the weight, and the 128-dimensional rockburst incubation stage prototype representation is updated.

[0119] The multi-source heterogeneous information fusion layer receives 6 sets of 128-dimensional stage prototype representations, which are spliced ​​together according to the time sequence of rockburst gestation to form a 768-dimensional serial vector. The network is configured with 128 fully connected neurons and a ReLU activation function. Based on the comparison between the normalized source response of the four types of monitoring data in each stage prototype and the third preset threshold, it adaptively retains the fusion pathway of high-contribution data and suppresses the low-contribution pathway. After fully connected transformation, it outputs a 128-dimensional rockburst gestation state fusion representation.

[0120] The output decision layer includes a dual-branch fully connected structure. The rockburst type branch is configured with a 4×128-dimensional weight matrix and a 4-dimensional bias, while the incubation stage branch is configured with a 6×128-dimensional weight matrix and a 6-dimensional bias. The two branches perform linear transformations on the fused representation of the rockburst incubation state and then normalize it using Softmax to obtain a decision response set. By taking the maximum response index and matching the corresponding rockburst type and incubation stage label, a one-to-one correspondence between type and stage is established, and the joint decision result of the two matching is output.

[0121] In one exemplary implementation, the process of reconstructing the energy stage constrained positive and negative affinity channels is as follows: Figure 4 As shown, it includes the following sub-steps:

[0122] S31. At the affinity channel generation position inside the unified prototype differential attention layer of the prototype differential attention neural network, the energy accumulation and release stage state is received, and the initial affinity relationship between the precursor monitoring variable and the prototype of the rockburst incubation stage generated by the unified prototype differential attention layer is obtained, forming the affinity relationship to be reconstructed as the object of positive and negative affinity channel reconstruction.

[0123] S32. Based on the energy accumulation and release stage state, extract the correspondence between elastic energy accumulation and microseismic and acoustic emission energy release at the affinity channel generation location, and map the correspondence to the precursor monitoring variables and rockburst incubation stage prototypes in the affinity relationship to be reconstructed, forming the initial affinity relationship of stage constraints.

[0124] S33. Based on the initial affinity relationship of the stage constraint, perform stage correspondence matching between each precursor monitoring variable and each rockburst incubation stage prototype, and calculate the stage correspondence value of the precursor monitoring variable relative to the rockburst incubation stage prototype to form a set of stage correspondence values.

[0125] S34. Compare the stage corresponding values ​​in the stage corresponding value set with the second preset threshold. If the stage corresponding value is not lower than the second preset threshold, generate a stage corresponding mark. If the stage corresponding value is lower than the second preset threshold, generate a stage non-corresponding mark to form an affinity polarity mark set.

[0126] S35. Based on the affinity polarity label set, configure positive affinity relationships between precursor monitoring variables with stage-corresponding labels in the initial affinity relationship of stage constraints and the prototype of the rockburst incubation stage, and configure negative affinity relationships between precursor monitoring variables with stage-non-corresponding labels and the prototype of the rockburst incubation stage. Update the affinity direction in the initial affinity relationship of stage constraints with positive and negative affinity relationships to form energy stage constraint affinity relationships.

[0127] S36. Based on the energy stage constraint affinity relationship, the affinity channels between the precursor monitoring variables and the rockburst incubation stage prototype in the unified prototype differential attention layer are separated into positive and negative channels to form energy stage constraint positive and negative affinity channels. This ensures that the precursor monitoring variables corresponding to the stage form a positive affinity relationship with the rockburst incubation stage prototype, and that the precursor monitoring variables not corresponding to the stage form a negative affinity relationship with the rockburst incubation stage prototype.

[0128] S4. Based on the energy stage constraint of positive and negative affinity channels, perform prototype attribution calculation on the monitoring variable sequence of rockburst incubation precursors to form a prototype representation of the rockburst incubation stage;

[0129] The purpose of this step is to determine the attribution of precursor variables to prototypes at each stage based on the constraint relationship of positive and negative affinity channels, aggregate effective monitoring features through positive attribution and negative exclusion differential operations, dynamically update the prototypes at each rockburst incubation stage, and make the prototype features fit the real energy evolution characteristics of the surrounding rock.

[0130] The core principle is as follows: When calculating the prototype attribution of each precursor monitoring variable based on the set of stage constraint attribution responses, firstly, the negative rejection response of each precursor monitoring variable relative to each rockburst incubation stage prototype is compared with the preset rejection conditions, and the rockburst incubation stage prototypes that meet the preset rejection conditions are excluded from the attribution calculation range of the corresponding precursor monitoring variable; then, among the rockburst incubation stage prototypes that have not been excluded, the rockburst incubation stage prototype to which the corresponding precursor monitoring variable belongs is determined based on the stage constraint attribution response; when multiple rockburst incubation stage prototypes meet the preset attribution conditions, the rockburst incubation stage prototype with the largest stage constraint attribution response is selected as the rockburst incubation stage prototype to which the corresponding precursor monitoring variable belongs, and the stage constraint prototype attribution relationship is formed accordingly.

[0131] In one exemplary implementation, the process for calculating the prototype representation of the rockburst incubation stage in this step is as follows: Figure 5 As shown, it includes the following sub-steps:

[0132] S41. Receive the positive and negative affinity channels of the energy stage constraint and the monitoring variable sequence of rockburst incubation precursors. At the affinity reading position of the prototype attribution calculation, pair the precursor monitoring variables in the rockburst incubation precursor monitoring variable sequence with the prototype of the rockburst incubation stage according to the sampling time to form a pairing relationship to be assigned.

[0133] S42. Based on the energy stage constraint of positive and negative affinity channels, perform channel reading on the positive and negative affinity relationships in the pairing relationships to be assigned, and write the reading results into the assignment determination criteria of the pairing relationships to be assigned, thus forming an energy-constrained pairing relationship to be assigned;

[0134] S43. Based on the energy-constrained pairing relationship, calculate the positive attribution response for precursor monitoring variables with positive affinity and the prototype of the rockburst incubation stage, and calculate the negative repulsion response for precursor monitoring variables with negative affinity and the prototype of the rockburst incubation stage, thus forming a positive and negative attribution response relationship.

[0135] S44. Based on the positive and negative attribution response relationship, the positive attribution response and negative rejection response of the same precursor monitoring variable relative to the same rockburst incubation stage prototype are calculated by difference to obtain the stage constraint attribution response of each precursor monitoring variable relative to each rockburst incubation stage prototype, forming a stage constraint attribution response set.

[0136] S45. Based on the set of stage constraint attribution responses, perform prototype attribution calculations for each precursor monitoring variable, assign precursor monitoring variables whose stage constraint attribution responses meet the preset attribution conditions to the corresponding rockburst incubation stage prototypes, and exclude precursor monitoring variables whose negative rejection responses meet the preset rejection conditions from the corresponding rockburst incubation stage prototypes, thus forming a stage constraint prototype attribution relationship.

[0137] S46. Based on the stage-constrained prototype attribution relationship, aggregate the precursor monitoring variables assigned to each rockburst incubation stage prototype, and limit the aggregation range of each rockburst incubation stage prototype with the stage-constrained prototype attribution relationship to form a prototype aggregation representation.

[0138] S47. Update the prototype of the rockburst incubation stage based on the prototype aggregation representation to form the prototype representation of the rockburst incubation stage.

[0139] S5. Based on the prototype representation of the rockburst incubation stage, multi-source heterogeneous information is fused to obtain the fused representation of the rockburst incubation state;

[0140] The purpose of this step is to adaptively retain or suppress the fusion pathway based on the updated stage prototype representation and the response intensity of various monitoring data at different stages, so as to achieve staged refined multi-source information fusion and output a unified rockburst incubation state fusion feature that takes into account the contributions of various monitoring signals.

[0141] In one exemplary implementation, the process for generating the fused representation of the rockburst incubation state in this step is as follows: Figure 6 As shown, it includes the following sub-steps:

[0142] S51. Receive the prototype representation of the rockburst incubation stage, determine the stage source based on the prototype representation of the rockburst incubation stage to which the precursor monitoring variables belong, and form a prototype representation of the stage source calibration.

[0143] S52. Based on the stage source calibration prototype representation, calculate the source response of microseismic data, acoustic emission data, stress data and displacement data in the same rockburst incubation stage prototype, and use the source response to update the fusion basis to form a source response constrained prototype representation;

[0144] S53. Based on the source response constraint prototype representation, compare the source response with the third preset threshold. If the source response is not lower than the third preset threshold, retain the intra-stage fusion path. If the source response is lower than the third preset threshold, suppress the intra-stage fusion path to form an intra-stage fusion representation.

[0145] S54. Based on the intra-stage fusion representation, multi-source heterogeneous information fusion is performed on the intra-stage fusion representations of different rockburst incubation stage prototypes, and the fusion order is constrained by the stage relationship between the rockburst incubation stage prototypes to obtain the fusion representation of the rockburst incubation state.

[0146] S6. Determine the rockburst type and incubation stage based on the fusion representation of rockburst incubation state, and output the fusion results of multi-source heterogeneous information of different types of rockburst incubation processes.

[0147] The purpose of this step is to complete the dual-dimensional determination of rockburst type and surrounding rock incubation stage based on the unified fusion characteristics of the incubation state, establish a one-to-one matching relationship between the two, and output complete rockburst multi-source information fusion early warning results, so as to provide direct decision-making basis for rockburst prevention and control in underground engineering.

[0148] In one exemplary implementation, the process for jointly determining the rockburst type and incubation stage in this step is as follows: Figure 7 As shown, it includes the following sub-steps:

[0149] S61. Receive the fused representation of rockburst incubation state, perform judgment response calculation on the fused representation of rockburst incubation state, and form a judgment response set corresponding to the rockburst type and incubation stage respectively;

[0150] S62. Based on the set of judgment responses, determine the location of the maximum response for the judgment response corresponding to the rockburst type to form a rockburst type judgment result, and determine the location of the maximum response for the judgment response corresponding to the incubation stage to form an incubation stage judgment result.

[0151] S63. Combine and match the results of rockburst type determination and incubation stage determination to form a joint determination result corresponding to rockburst type and incubation stage;

[0152] S64. Output the multi-source heterogeneous information fusion results of different types of rockburst incubation processes based on the joint judgment results.

[0153] Example:

[0154] This embodiment provides a method for fusing multi-source heterogeneous information on different types of rockburst incubation processes based on deep learning. The implementation process is as follows:

[0155] S1. Time-series alignment of multi-source monitoring data to construct a precursor variable sequence:

[0156] Acquire timestamped microseismic data, acoustic emission data, stress data, and displacement data within a time window of underground engineering excavation. Write the microseismic data as... Write the acoustic emission data as Write the stress data as Write the displacement data as .in, Indicates the source of microseismic data. Indicates the source of acoustic emission data. Indicates the source of stress data. Indicates the source of displacement data; , , and These represent the number of records from the four sources, respectively. , , and These represent the timestamps of the corresponding records. , , and These represent the monitoring field vectors for the corresponding records. The monitoring field vectors for microseismic data include microseismic energy, microseismic frequency, and microseismic event spatial density; for acoustic emission data, they include acoustic emission energy, acoustic emission count, and dominant acoustic emission frequency; for stress data, they include principal stress variation, tangential stress variation, and stress growth rate; and for displacement data, they include crown displacement, sidewall displacement, and displacement change rate. Records from the four sources are arranged in ascending order of acquisition timestamp.

[0157] The time window for underground engineering excavation is written as Using a preset sampling period Constructing candidate sampling time ,in , This represents the number of candidate sampling times, and each candidate sampling time... All are located within the underground engineering excavation time window. The time alignment window threshold is set to... Regarding the source ,according to Determine the source At the candidate sampling time The most recent record index, Indicates the index of the selected record; in In this case, the corresponding record will be used as the source. At the candidate sampling time Alignment records; when multiple records satisfy the same minimum time difference, select the record with the earliest sampling timestamp. Candidate sampling time. When simultaneously acquiring aligned records of microseismic data, acoustic emission data, stress data, and displacement data, candidate sampling times are retained. If candidate sampling times do not simultaneously obtain aligned records from all four sources, a sampling time monitoring variable group is not constructed.

[0158] The retained candidate sampling times are rewritten in ascending chronological order as follows: ,in , This indicates the number of sampling times to retain. The index of the most recent record, determined along with the retained candidate sampling times, will be synchronously rewritten. Sampling time Serial number within the underground engineering excavation time window and time value Together, determine the time location. Based on the time location, , , and Binding to the same sampling time To form a set of monitoring variables at the sampling time Monitoring variable group at sampling time The source boundaries of microseismic data, acoustic emission data, stress data, and displacement data are preserved. Monitoring fields from different sources are not averaged, and monitoring fields from different sampling times are not statistically merged.

[0159] Based on the sampling time, the source relationships of the monitored variable groups are set and their positions are fixed. The source arrangement rule is written as follows: , This represents the fixed arrangement order of microseismic data, acoustic emission data, stress data, and displacement data at each sampling time. The data is arranged according to their source relationships. To form an aligned monitoring variable set Align the monitoring variable groups The fields are ordered as follows: microseismic energy, microseismic frequency, microseismic event spatial density, acoustic emission energy, acoustic emission count, acoustic emission dominant frequency, principal stress variation, tangential stress variation, stress growth rate, crown displacement, sidewall displacement, and displacement variation rate. Monitoring fields from the same source occupy the same arrangement position at different sampling times, ensuring that microseismic data, acoustic emission data, stress data, and displacement data form a consistent field index within the underground engineering excavation time window.

[0160] The monitoring variable groups are connected and aligned according to the time sequence of the underground engineering excavation time window to form a monitoring variable sequence for precursors of rockburst incubation. ,in In one example, the underground engineering excavation time window contains continuous... At each valid sampling time, the sequence of monitoring variables for precursory rockbursts is as follows: The input mapping layer of the prototype differential attention neural network backbone is configured with the input mapping layer at each sampling time step. As input, the input mapping layer generates variable representations while maintaining a fixed arrangement of source relationships; the process of determining the state of the energy accumulation and release stage originates from the same... The system reads the fields corresponding to microseismic data, acoustic emission data, stress data, and displacement data, so that elastic energy accumulation, microseismic energy release, acoustic emission energy release, and displacement response changes have a unified sampling time index.

[0161] S2. Calculate the energy accumulation and release stage states, and quantify the temporal relationship of rock mass evolution:

[0162] Receive rockburst incubation precursor monitoring variable sequence , , This indicates the number of valid sampling moments within the underground engineering excavation time window. Each For aligned monitoring variable groups at the same sampling time, the positions are fixed according to their source relationships and then split into... , , and .symbol This represents a vector of microseismic data fields, containing microseismic energy, microseismic frequency, and spatial density of microseismic events; (symbol) This represents a vector of acoustic emission data fields, containing acoustic emission energy, acoustic emission count, and acoustic emission frequency; (symbol) This represents a vector of stress data fields, containing principal stress variations, tangential stress variations, and stress growth rate; (symbol) This represents a vector of displacement data fields, including crown displacement, sidewall displacement, and displacement rate of change. Following the temporal order of the underground engineering excavation time window, data from the same sampling time are... , , and Bind to time source variable group The source of the time variable group Maintain the source correspondence of microseismic data, acoustic emission data, stress data, and displacement data at the same sampling time.

[0163] Using preset stress field coefficients Processing stress data field vectors Preset stress field coefficient This represents the fixed participation ratio of principal stress variation, tangential stress variation, and stress growth rate in stress characterization, with preset stress field coefficients. Before the method is run, the three elements are determined by calibration samples or engineering settings, ensuring that none of them are negative and the sum of the three elements is one. The stress characterization value is obtained by multiplying the principal stress variation, tangential stress variation, and stress growth rate by their respective preset stress field coefficients and then summing them. .from Initially, the stress characterization value at the next adjacent sampling time is subtracted from the stress characterization value at the next adjacent sampling time to obtain the stress variation between adjacent sampling times. ,exist Mark adjacent stress changes as invalid. (This refers to marking adjacent stress changes as invalid.) Compared with the preset stress growth threshold Compare; when Not lower than the preset stress growth threshold At that time, the direction of stress growth is positive, and the magnitude of stress growth is [value missing]. The absolute amplitude, the duration of stress increase is the product of the number of consecutive positive sampling moments and the sampling period; when Below the preset stress growth threshold When the stress growth direction is non-positive, the duration of stress growth is recounted.

[0164] Mapping the direction of stress growth, the magnitude of stress growth, and the duration of stress growth into an elastic energy accumulation benchmark. Elastic energy accumulation benchmark It consists of a stress growth direction marker, stress growth amplitude, stress duration, and a valid marker; when the stress growth direction is positive and the stress growth amplitude is not lower than the preset stress growth threshold. The duration of continuous stress increase is not less than a preset duration threshold. When the above conditions are not met simultaneously, the valid marker is 1; when the above conditions are not met simultaneously, the valid marker is 0. The elastic energy accumulation benchmark is arranged according to the sampling time. Forming a candidate sequence of stage states .

[0165] Based on the candidate sequence of stage states A continuous and effective elastic energy accumulation benchmark is used to determine the sampling time interval for continuous increase in stress data. , , Indicates the number of sampling time intervals. Indicates the starting index of the sampling time interval. Indicates the end index of the sampling time interval. Preset number of adjacent sampling times. Used to define the matching range for energy release, index set Indicates the sampling time interval The subsequent adjacent sampling times. The sampling time intervals in which stress data continues to increase are used to define the sequential boundaries of the elastic energy accumulation benchmark on the time axis.

[0166] Using preset microseismic field coefficients Processing microseismic data field vectors Preset microseismic field coefficients This represents the fixed participation ratio of microseismic energy, microseismic frequency, and microseismic event spatial density in the microseismic release characterization. The microseismic release characterization value is obtained by multiplying microseismic energy, microseismic frequency, and microseismic event spatial density by their respective preset microseismic field coefficients and then summing them. .from Initially, the microseismic energy release variation is generated based on the changes in microseismic release characterization values ​​between adjacent sampling times. ,exist Changes in microseismic energy release are marked as invalid. Not lower than the preset micro-seismic release threshold At that time, the sampling time will be This is recorded as the sampling time for microseismic energy release. If the sampling time for microseismic energy release is earlier than the starting index of the entire sampling time interval, it is determined that the change in microseismic energy release occurs before the elastic energy accumulation reference, and is within the sampling time... The writing phase does not correspond; the sampling time for microseismic energy release is located at any... Within the timeframe, the changes in microseismic energy release were successfully matched with the corresponding elastic energy accumulation benchmark, and this was achieved in the candidate sequence of stage states. The corresponding sampling time is written into the microseismic matching marker to form a candidate sequence of microseismic constraint stage states. .

[0167] Using preset acoustic emission field coefficients Processing acoustic emission data field vectors Preset acoustic emission field coefficients This represents the fixed participation ratio of acoustic emission energy, acoustic emission count, and acoustic emission dominant frequency in the acoustic emission release characterization. The acoustic emission release characterization value is obtained by multiplying the acoustic emission energy, acoustic emission count, and acoustic emission dominant frequency by their respective preset acoustic emission field coefficients and then summing them. .from Initially, the acoustic emission energy release variation is generated based on the changes in acoustic emission release characterization values ​​between adjacent sampling times. ,exist The change in acoustic emission energy release is marked as invalid. Not lower than the preset acoustic emission release threshold At that time, the sampling time will be This is denoted as the acoustic emission energy release sampling time. If the acoustic emission energy release sampling time is earlier than the starting index of the entire sampling time interval, then at that sampling time... The writing phase does not correspond to the previous one; the acoustic emission energy release sampling time is located at any time. During this period, the changes in acoustic emission energy release were correlated with the candidate state sequence of the microseismic constraint stage. The elastic energy accumulation benchmark was successfully matched, and an acoustic emission matching marker was written at the corresponding sampling time to form a candidate sequence of states for the microseismic and acoustic emission constraint stages. .

[0168] Using preset displacement field coefficients Processing displacement data field vectors Preset displacement field coefficients This represents the fixed participation proportion of crown displacement, sidewall displacement, and displacement rate of change in the displacement response characterization. The displacement response characterization value is obtained by multiplying the crown displacement, sidewall displacement, and displacement rate of change by their respective preset displacement field coefficients and then summing them. .from Initially, displacement response changes are generated based on the changes in displacement response characterization values ​​between adjacent sampling times. ,exist The displacement response change is marked as invalid. Not lower than the preset displacement response threshold At that time, the sampling time will be The effective sampling time of the displacement response is denoted as [time]. The effective sampling time of the displacement response is then compared with the candidate state sequences of the microseismic and acoustic emission constraint stages. Consistency of sampling times is determined at the same sampling time. When the same sampling time has microseismic matching markers, acoustic emission matching markers, and valid displacement response markers, the corresponding sampling time is retained as a stage-corresponding candidate. If any marker is missing, the corresponding sampling time is updated to a stage-non-corresponding candidate, forming a displacement constraint stage state candidate sequence. .

[0169] Candidate sequence of states for displacement constraint stage Calibration results are generated at each sampling time. Calibration results It includes four criteria: effective elastic energy accumulation benchmark, microseismic energy release change after the elastic energy accumulation benchmark, acoustic emission energy release change after the elastic energy accumulation benchmark, and meeting the preset displacement response threshold. The displacement response change. First preset threshold condition. Based on the preset stress growth threshold Preset duration threshold Preset micro-vibration release threshold Preset acoustic emission release threshold Preset displacement response threshold and the number of adjacent sampling times Common constraints. The calibration results will be used for calibration. With the first preset threshold condition A comparison is made; when all four criteria meet the first preset threshold condition... At that time, the sampling time will be Marked as stage correspondence When at least one judgment item does not meet the first preset threshold condition At that time, the sampling time will be Marked as a stage mismatch relationship Arranged according to sampling time This forms a sequence corresponding to the accumulation and release of energy. .

[0170] According to the corresponding sequence of energy accumulation and release Generation of energy accumulation and release phase state , Quantity Divide the number of sampling times corresponding to the effective elastic energy accumulation benchmark by Receive, Quantity Divide the number of microseismic energy release sampling times corresponding to the elastic energy accumulation reference formation time by Receive, component Divide the number of acoustic emission energy release sampling times corresponding to the elastic energy accumulation reference formation time by Receive, component The number of valid sampling times of displacement response occurring synchronously with the stage correspondence is divided by Received. In one example, The monitoring variable sequence for precursors of rockburst formation is as follows: The input sequence, the energy accumulation and release phase states This is a four-dimensional state vector, representing the states during the energy accumulation and release phase. The affinity channel generation location of the unified prototype differential attention layer in the prototype differential attention neural network is used to represent the correspondence between elastic energy accumulation and microseismic and acoustic emission energy release.

[0171] S3. Reconstruct the positive and negative affinity channels of the unified prototype differential attention layer to differentiate stage-matched variables:

[0172] Receive rockburst incubation precursor monitoring variable sequence and the energy accumulation and release phase state ,in , This indicates the number of valid sampling moments within the underground engineering excavation time window. Each This represents a precursor monitoring variable at a sampling time. Representing the real number field, the twelve fields are arranged according to microseismic energy, microseismic frequency, spatial density of microseismic events, acoustic emission energy, acoustic emission count, dominant acoustic emission frequency, principal stress variation, tangential stress variation, stress growth rate, crown displacement, sidewall displacement, and displacement variation rate. Each Indicates the sampling time The energy accumulation and release phase, in which Corresponding to the accumulation of elastic energy, Corresponding to the release of energy from microseismic events. Corresponding to the release of acoustic emission energy. The corresponding displacement response changes. The components in the energy accumulation and release phase are dimensionless normalized values, with a range of [value missing]. This is used to generate the affinity channel for entering the unified prototype differential attention layer.

[0173] The prototype differential attention neural network first processes the sequence of monitoring variables for precursors of rockburst formation through an input mapping layer. The input mapping layer is denoted as... , This represents the weights and biases of the input mapping layer. The input mapping layer consists of 64 fully connected neurons, each receiving... The 12 input connections utilize linear rectifier units for nonlinear units. Each... enter Generate 64-dimensional variable representation The time window coding layer is denoted as... , This represents the weights and biases of the temporal window coding layer. The temporal window coding layer consists of 128 fully connected neurons, each receiving a fully connected input represented by a 64-dimensional variable. enter Generate a 128-dimensional time representation .

[0174] Unified Prototype Differential Attention Layer Reception Six prototype stages of rockburst incubation were set up. ,in This represents a set of prototypes for the rockburst incubation stage. Indicates the prototype index of the rockburst incubation stage, each This represents a prototype of a rockburst incubation stage. The unified prototype differential attention layer, at the affinity channel generation location, for each... With each Fully connected computation is performed, where affinity calculation results are generated jointly from the components of the 128-dimensional time representation, the components of the 128-dimensional rockburst incubation stage prototypes, the corresponding connection weights, and biases. The unified prototype differential attention layer uses normalization units to compress the affinity calculation results corresponding to the six rockburst incubation stage prototypes at the same sampling time to a single unit. To form an initial affinity relationship Initial affinity Indicates the sampling time Precursor monitoring variables and the first Data proximity relationships between prototypes of the rockburst incubation stage. Affinity channels are generated within the differential attention layer of the unified prototype. The calculation location is determined before the prototype attribution calculation. As the object of positive and negative affinity channel reconstruction, it forms an affinity relationship to be reconstructed. Affinity relationships to be reconstructed The affinity strength field and same.

[0175] Receive energy accumulation and release phase state at the same affinity channel generation location The affinity channel reconstruction calculation is denoted as... , This represents the weights and biases in the affinity channel reconstruction computation. The affinity channel reconstruction computation involves six fully connected neurons, each receiving the 4-dimensional energy accumulation and release phase states. The nonlinear units are hyperbolic tangent units. enter Generate 6-dimensional stage constraint results .in, Indicates the sampling time The corresponding stage constraint result vectors for the six prototype rockburst incubation stages. Indicates the sampling time The correspondence between elastic energy accumulation and micro-vibrations and acoustic emission energy release is relative to the first The constraint results for the prototype of the rockburst incubation stage are within the range of [value range missing]. .Will Write the affinity relationship to be reconstructed The stage constraint position forms the initial affinity relationship of the stage constraint. Phase constraints on initial affinity It includes an affinity strength field and a stage constraint field. The affinity strength field is derived from the affinity relationship to be reconstructed. The stage constraint field is derived from the stage constraint result. .

[0176] Initial affinity based on phase constraints For each sampling time Precursor monitoring variables and prototypes of each rockburst incubation stage Perform stage-to-stage matching. Stage-to-stage matching read. The initial affinity strength, read The phase constraint direction is determined, and the energy accumulation and release phase state is read. The four underlying components are used. To ensure that the stage correspondence matching simultaneously reflects the data proximity relationship, the correspondence between elastic energy accumulation and microseismic energy release, the correspondence between elastic energy accumulation and acoustic emission energy release, and the displacement response change, the stage correspondence values ​​of the precursor monitoring variables relative to the prototype of the rockburst incubation stage are calculated in the following way:

[0177] ;

[0178] in, Indicates the sampling time Precursor monitoring variables relative to the first The corresponding stage values ​​of a prototype rockburst incubation stage; Indicates the sampling time index; Indicates the prototype index of the rockburst incubation stage; Indicates the sampling time Precursor monitoring variables and the first The initial affinity between the prototypes of the rockburst incubation stage; Indicates the sampling time The energy accumulation and release phase state relative to the first Stage constraint results of a prototype rockburst incubation stage; Indicates the sampling time The elastic energy accumulation component; Indicates the sampling time The energy release component of microseismic events; Indicates the sampling time The component of acoustic emission energy released; Indicates the sampling time The displacement response change component; This indicates that the smaller of the two input values ​​is selected. Represents absolute value operations; numerical values Indicates the upper limit of normalization; numerical value Indicates will from Switch to The normalization scale used; numerical This indicates the number of items corresponding to elastic energy accumulation and microseismic energy release, elastic energy accumulation and acoustic emission energy release, and displacement response changes. According to... and index arrangement The set of values ​​corresponding to the formation stage The set of values ​​corresponding to each stage The size is .

[0179] Set the threshold corresponding to the preset stage. .when Not lower than the threshold corresponding to the preset stage At that time, the corresponding marker for the generation stage ;when Below the preset threshold At that time, the generation stage does not correspond to the tag. .in, Indicates the sampling time Precursor monitoring variables and the first Affinity polarity markers between prototypes of the rockburst incubation stage. According to and index arrangement This forms a set of affinity polarity markers. Affinity polarity tag set The size is .

[0180] Based on the set of affinity polarity markers The update phase constrains the initial affinity. Affinity direction in the matrix. Set a preset positive threshold. When the initial affinity Meets the preset positive threshold ,and When the stage does not correspond to the marker, the sampling time will be... Precursor monitoring variables and the first The affinity direction between the prototypes of the rockburst incubation stage is updated to a negative affinity; when When marking the stage, the sampling time will be... Precursor monitoring variables and the first The affinity directions between prototypes in the rockburst incubation stage are updated to positive affinity. Both positive and negative affinity are generated within the unified prototype differential attention layer, without adding rule judgments after the decision results of the prototype differential attention neural network. The prototypes are then arranged according to the updated affinity directions. Forming energy stage-constrained affinity relationships Energy stage constrains affinity relationships. It includes an affinity strength field and an affinity direction field. The affinity strength field is derived from the initial affinity relationship of the stage constraint. The affinity strength field and the affinity direction field are composed of a set of affinity polarity markers. Updated.

[0181] Affinity is constrained by energy stage. Positive and negative channels are separated in the affinity channels of the unified prototype differential attention layer. Channels with positive affinity are... Write to positive affinity channel And set the negative affinity channels at the same index position to zero; and set the channels with negative affinity to zero. Write to negative affinity channel And the positive affinity channels at the same index position are set to zero. This forms an energy stage constraint for positive and negative affinity channels. In one example, The monitoring variable sequence for precursors of rockburst formation is as follows: Time is expressed as The initial affinity is The stage constraint result is The set of values ​​corresponding to each stage is The set of affinity polarity markers is Both positive affinity channels and negative affinity channels are .

[0182] S4. Perform prototype attribution differential calculation and update the prototype representation of the rockburst incubation stage:

[0183] Energy receiving stage constrains positive and negative affinity channels And rockburst precursor monitoring variable sequence .in, Indicates a positive affinity channel. Indicates a negative affinity channel. and All dimensions are ; This indicates the number of valid sampling moments within the underground engineering excavation time window; Indicates the sampling time index; Indicates the sampling time Precursor monitoring variables, Representing the real number field, the twelve fields are arranged according to microseismic energy, microseismic frequency, microseismic event spatial density, acoustic emission energy, acoustic emission count, acoustic emission dominant frequency, principal stress variation, tangential stress variation, stress growth rate, crown displacement, sidewall displacement, and displacement change rate. Positive affinity channel. Indicates the sampling time Precursor monitoring variables and the first Positive affinity strength and negative affinity channels between prototypes of the rockburst incubation stage. Indicates the sampling time Precursor monitoring variables and the first The negative affinity strength between prototypes in the rockburst incubation stage This represents the prototype index for the rockburst incubation stage. Both positive and negative affinity intensities are dimensionless values ​​output from the unified prototype differential attention layer.

[0184] In the prototype differential attention neural network, the prototype attribution computation layer follows the unified prototype differential attention layer. The prototype attribution computation layer reads the temporal representation generated by the input mapping layer and the temporal window encoding layer. The input mapping layer is denoted as... , This represents the weights and biases of the input mapping layer, which consists of 64 fully connected neurons, each receiving... It has 12 input connections, and the nonlinear units use linear rectified units. The time window coding layer is denoted as... , This represents the weights and biases of the temporal window coding layer, which consists of 128 fully connected neurons, each receiving a fully connected input represented by a 64-dimensional variable. enter Generate 64-dimensional variable representation Then enter Generate a 128-dimensional time representation Time representation As a vector representation of precursor monitoring variables in prototype attribution calculation. All 128 components are dimensionless values ​​at the output of the time window coding layer.

[0185] At the affinity read position in the prototype attribution calculation, the precursor monitoring variables in the rockburst incubation precursor monitoring variable sequence are paired with the prototypes of the rockburst incubation stage according to the sampling time. The six prototypes of the rockburst incubation stage are written as follows: ,in This represents a set of prototypes for the rockburst incubation stage, each of which... This represents a prototype of a rockburst gestation stage. (Rockburst gestation stage prototype) To unify the prototype parameters in the prototype differential attention layer, they are determined during the training process and stored in the unified prototype differential attention layer before the prototype differential attention neural network is executed. All 128 components are related to time representation. Dimensionless numerical values ​​located in the same prototype space. For each sampling time... and a prototype index for each rockburst incubation stage Constructing matching terms Pairing items Includes sampling time index Precursor monitoring variables Time representation Prototype of the rockburst incubation stage Prototype Index of Rockburst Incubation Stage .according to and The index arranges all paired items, forming a pairing relationship to be assigned. Pairing relationships to be assigned The size is .

[0186] Based on the energy stage constraint of positive and negative affinity channels, the attribution pairing relationship is considered. Each pair is read from a channel. The affinity read positions for prototype attribution calculation follow the same... and Index Reading and The positive and negative affinity strengths read are then written into the pairing terms. The field used to determine ownership. The paired item after writing is denoted as... After all the pairs are written, an energy-constrained pairing relationship is formed. Energy-constrained pairing relationships to be assigned. It includes time representation, prototype of rockburst incubation stage, positive affinity intensity, and negative affinity intensity.

[0187] Based on the energy constraint of the pairing relationship to be assigned For precursor monitoring variables with positive affinity and prototypes in the rockburst incubation stage, positive attribution responses are calculated. For precursor monitoring variables with negative affinity and prototypes in the rockburst incubation stage, negative repulsion responses were calculated. .when When it is greater than zero, As a positive attribution response ;when When it equals zero, Set to zero. When When it is greater than zero, As a negative repulsion response ;when When it equals zero, Set to zero. Sort by the same index. and This forms a positive and negative attribution response relationship.

[0188] Based on the positive and negative attribution response relationship, the positive attribution response and negative repulsion response of the same precursor monitoring variable relative to the same rockburst incubation stage prototype are differentially processed. This differential processing focuses on the stage-constrained attribution response, using positive affinity channel readouts, negative affinity channel readouts, time representation components, and rockburst incubation stage prototype components to jointly limit the prototype attribution intensity. This ensures that the stage-constrained attribution response simultaneously embodies the correspondence between the energy stage constraint positive and negative affinity channels and the prototype spatial components. Specifically:

[0189] ;

[0190] in, Indicates the sampling time Precursor monitoring variables relative to the first Stage constraint attribution response of a prototype rockburst incubation stage; Indicates the sampling time Precursor monitoring variables relative to the first Positive attribution response of a prototype of a rockburst incubation stage; Indicates the sampling time Precursor monitoring variables relative to the first The negative repulsion response of a prototype rockburst incubation stage; Indicates the sampling time index; Indicates the prototype index of the rockburst incubation stage; Component indexes representing the time representation and prototype of the rockburst incubation stage; Indicates the sampling time 128-dimensional time representation The Middle One output component; Indicates the first A prototype of the rockburst incubation stage The Middle One prototype component; This indicates that the smaller of the two input values ​​is selected. This indicates that the larger of the two input values ​​is selected. This represents the absolute value operation; This indicates summing over 128 components; This indicates a pre-defined positive stable term. Set as before the method runs. This is used to avoid the denominator being zero. In the formula... , , and All values ​​are directly readable; the fractional terms consist of dimensionless values ​​within the same prototype space; and the stage constraint belongs to the response. This is a dimensionless numerical value. According to... and The indexes are arranged to assign all stage constraint responses, forming a set of stage constraint response assignments. Stage constraint attribution response set The size is .

[0191] According to the stage constraint, the response set is assigned. For each precursor monitoring variable, perform prototype attribution calculations. Set preset exclusion conditions. and preset attribution conditions ,in This represents the threshold for negative repulsion response. Indicates the threshold for attribution of stage constraints. and All parameters are determined by training sample calibration or engineering settings before the prototype differential attention neural network is executed. For the sampling time... First, put each one Rejection conditions Compare; when Not lower than the preset rejection condition At that time, the first A prototype of the rockburst incubation stage from the sampling time Excluded from the scope of the attribution calculation. Then, in the prototype of the rockburst incubation stage that was not excluded, With preset attribution conditions Comparison; when a prototype of a rockburst incubation stage meets the preset attribution conditions. At that time, the sampling time will be The precursor monitoring variables are assigned to the corresponding rockburst incubation stage prototypes; when multiple rockburst incubation stage prototypes meet the preset assignment conditions... At that time, the prototype of the rockburst incubation stage with the largest stage constraint attribution response was selected as the sampling time. The precursor monitoring variables are attributed to the prototype of the rockburst incubation stage; when the prototype of the unexcluded rockburst incubation stage does not exist, the sampling time is... The precursor monitoring variables are not included in the aggregation range of any rockburst incubation stage prototype; when none of the unexcluded rockburst incubation stage prototypes meet the preset attribution conditions. At that time, sampling time Precursor monitoring variables are not included in the stage constraint prototype attribution relationship. The stage constraint prototype attribution relationship is formed based on the above attribution results. , Indicates the sampling time The precursor monitoring variables are attributed to the prototype index of the rockburst incubation stage.

[0192] Based on the stage constraint prototype attribution relationship The precursor monitoring variables assigned to each prototype of the rockburst incubation stage were aggregated. For the first... A prototype of a rockburst incubation stage, constructing an allocation index set. ,in Indicates belonging to the first A set of sampling time indices for each prototype of the rockburst incubation stage. 128-dimensional time representation corresponding to each sampling time point As an aggregate object, and using stage constraints to define prototype attribution relationships. Limited to the first The aggregation range of the prototype of the rockburst incubation stage. The non-negative stage constraint attribution responses at each sampling time are used as the source of aggregation weights. The sum of non-negative stage constraint attribution responses within the same rockburst incubation stage prototype is used as the normalization divisor. When the normalization divisor is greater than zero, each non-negative stage constraint attribution response is divided by the normalization divisor to obtain the aggregation weights. These weights are then applied dimension-wise to the corresponding 128-dimensional time representation and summed sequentially to form the prototype aggregation representation. When the normalized divisor is equal to zero, for The 128-dimensional time representations corresponding to each sampling time point are summed dimension by dimension with the same weights to form a prototype aggregated representation. .when When empty, the prototype of the rockburst incubation stage will be displayed. As a prototype aggregation representation Each prototype aggregation represents .

[0193] Update the rockburst incubation stage prototype based on the prototype aggregation representation. (The remaining text appears to be incomplete and requires further context.) A prototype aggregation representation of the rockburst incubation stage. Write to the The representation of the prototype of each rockburst incubation stage forms the prototype representation of the rockburst incubation stage. The prototype representation of all rockburst incubation stages is denoted as... Each of them In one example, The energy stage constrains both the positive and negative affinity channels. The set of responses to stage constraints is The prototype of the rockburst incubation stage is represented by 6 128-dimensional vectors.

[0194] S5. Staged adaptive multi-source information fusion to generate a fusion representation of the overall gestation state:

[0195] Receiving the prototype representation of the rockburst incubation stage And read the prototype representation of the rockburst incubation stage. The same index stores the set of allocation indexes Simultaneously read the monitoring variable sequence of precursors to rockburst incubation. The twelve fields corresponding to the sampling time. Symbols Indicates the prototype index of the rockburst incubation stage, each Indicates the first A 128-dimensional representation of the prototype of the rockburst incubation stage. Represents the real number field. The 128 components correspond to the 128 output channels of the temporal window encoding layer in the prototype differential attention neural network. (Symbol) Indicates belonging to the first A set of sampling time indices for precursor monitoring variables of a rockburst incubation stage prototype. Symbols Indicates the number of valid sampling moments within the underground engineering excavation time window, symbol Indicates the sampling time Precursor monitoring variables. Written as The twelve fields are, in order: microseismic energy, microseismic frequency, microseismic event spatial density, acoustic emission energy, acoustic emission count, acoustic emission dominant frequency, principal stress variation, tangential stress variation, stress growth rate, crown displacement, sidewall displacement, and displacement variation rate.

[0196] Set up a phase relationship table ,symbol The symbols represent the stage relationships between prototypes in the rockburst incubation stage. Indicates the first Each rockburst incubation stage prototype corresponds to a specific stage origin. Stage Relationship Table The rockburst incubation stage prototypes are set according to their chronological relationship, and maintain the same index order as the six rockburst incubation stage prototypes in the unified prototype differential attention layer. In one example, to The stages are sequentially set as: initial accumulation stage, continuous accumulation stage, release initiation stage, release enhancement stage, instability imminent stage, and instability release stage. (Based on the allocation index set) It will belong to the first The precursor monitoring variables of the prototype rockburst incubation stage are labeled as the stage source. .when If empty, retain the first... The prototype of the rockburst incubation stage corresponds to the stage source. Source of the stage Write the corresponding Formation stage source identification prototype representation Stage source calibration prototype representation reserve The 128-dimensional prototype is represented, and the stage source is added. As stage calibration information.

[0197] Based on the stage source, the prototype is represented. Calculate the source response of microseismic data, acoustic emission data, stress data, and displacement data in the same prototype during the same rockburst incubation stage. Write the field index set as... , , , .symbol Corresponding microseismic data field, symbol Corresponding acoustic emission data field, symbol Corresponding stress data field, symbol Corresponding displacement data fields. Index each field. Set preset field lower limit and preset field limit and make Greater than .when Not higher than At that time, normalize the field values. Set to 0; when Not less than At that time, normalize the field values. Set to 1; when lie in and In between, minus The difference obtained afterwards is divided by and The difference is used to obtain the normalized field value. Normalized field values It is a dimensionless numerical value, and its range is [value range missing]. .

[0198] The source response calculation location is determined by using the intensity of the underlying field belonging to the same rockburst incubation stage prototype and the changes in adjacent sampling times.

[0199] ;

[0200] symbol Indicates the first Class source in the Source response in a prototype rockburst incubation stage; symbol Indicates the source type of the current calculation. ;symbol Indicates the source of microseismic data; symbol Indicates the source of acoustic emission data; symbol Indicates the source of stress data; symbol Indicates the source of displacement data; symbol Indicates the prototype index of the rockburst incubation stage; symbol Indicates belonging to the first A set of indexes of sampling times for precursor monitoring variables in the prototype stage of rockburst incubation; symbols Indicates the sampling time index; symbol Indicates a field index; symbol Indicates the first The set of field indexes corresponding to the class source; symbols Indicates the sampling time The Normalized values ​​for each field; symbol Indicates the sampling time The first adjacent sampling time Normalized values ​​for each field, when At that time, Set as ;symbol This represents the source enumeration variable used to iterate through the four types of sources; symbol Indicates the first The set of field indexes corresponding to the class source; symbols Represents absolute value operation; symbol This indicates a summation operation; the value 3 indicates that each source category contains three fields; the value 1 indicates the baseline bias of the change term between adjacent sampling times; [Symbol] This indicates a pre-defined positive stable term. Pick This is used to avoid the denominator being zero. All values ​​involved in the calculation in the formula are dimensionless. When empty, , , and All values ​​are set to zero. Sources are arranged according to the four source categories to form source-response vectors. .

[0201] Source response vector Write the fusion basis .symbol Indicates the first The basis for merging prototypes of the rockburst incubation stage. Basis for merging. It consists of four source path weights and four source path states. The four source paths are the microseismic data stage fusion path, the acoustic emission data stage fusion path, the stress data stage fusion path, and the displacement data stage fusion path. The source response vector is then written. Subsequently, the path weights of the fusion pathways within the microseismic data phase are: The path weights of the fusion pathways within the acoustic emission data phase are: The path weights of the fusion pathways within the stress data phase are: The path weights of the fusion pathways within the displacement data stage are: The fusion basis Representation of the prototype with stage source calibration Binding, forming a prototype representation of source-response constraints .

[0202] Set preset source response threshold ,symbol This represents the threshold used to determine whether the fusion pathway is preserved or suppressed within a given stage, with a preset source response threshold. The threshold is determined by training sample calibration or engineering settings before execution of the prototype differential attention neural network. In one example, a preset source response threshold is used. Set to 0.25. Based on the source response constraint prototype representation. , source response vector Each component in the data is related to a preset source response threshold. Comparison. When Not lower than the preset source response threshold At that time, the first one is reserved. A fusion pathway for microseismic data within the prototype rockburst incubation stage; when Below the preset source response threshold At that time, suppress the first The microseismic data fusion pathway within the prototype rockburst incubation stage is used. The fusion pathways within the acoustic emission data stage, stress data stage, and displacement data stage are preserved or suppressed according to the same threshold comparison rule.

[0203] For the A prototype of the rockburst incubation stage, representing the source calibration prototype of the fusion pathway receiving stage within the preserved stage. And according to the corresponding source response components The 128 components are weighted dimension-wise; the suppressed intra-stage fusion pathway outputs a 128-dimensional zero vector. The outputs of the retained intra-stage fusion pathways are summed dimension-wise and normalized according to the sum of the source responses of the retained pathways to form the intra-stage fusion representation. .symbol Indicates the first The intra-stage fusion representation corresponding to the prototype of the rockburst incubation stage. When the first... When the fusion pathways within all four stages of the prototype rockburst incubation stage are suppressed, Set to a 128-dimensional zero vector.

[0204] According to the stage relationship table The fusion representations within the six stages are sorted. These sorted fusion representations are then connected according to the chronological relationship of the gestation stages, forming a 768-dimensional connected input. The multi-source heterogeneous information fusion layer comprises 128 fully connected neurons, each receiving the 768-dimensional connected input from the fusion representations within the six stages. Each connected input includes weights and biases, and the nonlinear units are linear rectified units. The multi-source heterogeneous information fusion layer performs a fully connected transformation on the 768-dimensional connected input to form a fusion representation of the rockburst gestation state. ,symbol This represents a 128-dimensional fusion representation of the rockburst incubation state. In one example, the prototype representation of the rockburst incubation stage is six 128-dimensional vectors, the fusion representation within the stage is also six 128-dimensional vectors, the concatenation input is 768-dimensional, and the fusion representation of the rockburst incubation state is 128-dimensional.

[0205] S6. Jointly determine the rockburst type and incubation stage, and output the fused early warning results:

[0206] Receive rockburst incubation state fusion representation ,in, , Represents the real number field. Indicates the fusion component index, The fusion of the rockburst incubation state is represented by the first... Each fusion component. The first layer in the multi-source heterogeneous information fusion layer The output values ​​of each fully connected neuron are fused from the intra-stage fusion representation inputs corresponding to the six prototype rockburst gestation stages. Rockburst gestation state fusion representation. It serves as the input for the output layer of the prototype differential attention neural network to perform decision response calculations.

[0207] Set the output layer ,in, This represents the output layer parameters, which include the output weight for rockburst type. Rockburst type output bias Output weights during the gestation stage Output bias during the gestation stage Rockburst type output weight Rockburst type output bias Output weights during the gestation stage Output bias during the gestation stage The aforementioned weights and biases are determined during the training process and stored in the output layer before the prototype differential attention neural network is executed.

[0208] The output layer consists of 10 fully connected neurons. The first four fully connected neurons correspond to the rockburst type, and the last six fully connected neurons correspond to the gestation stage. The output position for the rockburst type is set... This indicates the output index for the rockburst type; the output layer will have 128 fused components. Multiply by respectively The Middle The 128 weights corresponding to each row are then multiplied by the 128 products. The Middle Adding the biases together yields the linear response of the rockburst type. Set the output location for the gestation stage. This indicates the output index during the incubation stage; the output layer will have 128 fused components. Multiply by respectively The Middle The 128 weights corresponding to each row are then multiplied by the 128 products. The Middle Adding the biases together yields the linear response during the gestation stage. .

[0209] The linear responses of the four rockburst types were processed using normalized exponential units. The normalized exponential units were respectively used for... , , and The natural index is taken, and the sum of the responses of the four natural indices is used as the normalization divisor to obtain the rockburst type determination response. .in, Indicates the response for determining the type of rockburst. Indicates the first The response for determining the location of each rockburst type is calculated. The linear responses for the six incubation stages are processed using normalized exponential units to obtain the incubation stage determination responses. .in, Indicates the response during the gestation stage. Indicates the first The response determines the output position for each gestation stage. and Arranged in order of output layer, forming a decision response set. .

[0210] Set up a rockburst type label table ,in, A table showing rockburst type labels. , , and These represent four preset rockburst types. In one example, the four preset rockburst types are, in order, strain-induced rockburst, tectonic rockburst, impact-induced rockburst, and composite rockburst. A label table for the incubation stage is set. ,in, Labels indicating the stages of pregnancy to These stages, in sequence, represent the initial accumulation stage, the continuous accumulation stage, the release initiation stage, the release enhancement stage, the instability imminent stage, and the instability release stage.

[0211] Determine the response based on the type of rockburst. The maximum response location was determined for the output locations of the four rockburst types. , , and The position of the output with the largest median value is denoted as ,in, This indicates the location of the maximum response for a particular rockburst type. When multiple rockburst types have the same maximum response value at their output locations, the rockburst type label table will be used. The output position with the smaller index is determined .Will As a result of rockburst type determination ,in, This indicates the rockburst type determined by the prototype differential attention neural network.

[0212] Response determined based on gestation stage The maximum response position is determined for the output positions of the six gestation stages. to The position of the output with the largest median value is denoted as ,in, This indicates the location of the maximum response during the gestation stage. When multiple gestation stage output locations have the same maximum response value, the gestation stage label table will be used. The output position with the smaller index is determined .Will As a result of determining the gestation stage ,in, This represents the gestation stage determined by the prototype differential attention neural network.

[0213] Based on the rockburst type determination results and gestation stage determination results Perform combined matching. Combined matching does not change the rockburst type determination response. and gestation stage determination response The numerical order will determine the rockburst type. Results of gestation stage determination Establish corresponding relationships within the same underground engineering excavation time window to form a joint judgment result. .in, This indicates the correspondence between rockburst type and gestation stage. Based on the joint assessment results... This forms the result of rockburst type determination. Result of determination of gestation stage And the correspondence between rockburst types and gestation stages The prototype differential attention neural network performs one forward inference for a single underground engineering excavation time window; for continuous underground engineering excavation time windows, the corresponding rockburst incubation state representations are input sequentially according to the time window order. The corresponding rockburst type determination results, gestation stage determination results, and the correspondence between rockburst type and gestation stage are generated respectively.

[0214] See Figure 8 This invention discloses the process by introducing energy accumulation and release stage constraints into a prototype differential attention neural network in the intermediate layer, and completing the prototype differential calculation, attribution calculation, and fusion representation of the rockburst incubation state. The left input includes four data streams: stress, microseismic data, acoustic emission data, and displacement data. Stress is used to determine the elastic energy accumulation baseline, while microseismic, acoustic emission, and displacement data are used to calculate corresponding changes and form stage states. After the stage states enter the prototype differential attention neural network, they participate in the prototype differential calculation together with precursor monitoring variables and the prototype of the rockburst incubation stage, forming an affinity matrix. The affinity matrix is ​​then reconstructed into positive and negative channels, with positive channels corresponding to positive affinity and negative channels corresponding to negative affinity. The attribution calculation part performs differential calculation based on positive attribution and negative repulsion, forming prototype attribution and prototype representation. The right fusion path retains or suppresses the microseismic response, acoustic emission response, stress response, and displacement response, forming a fusion representation of the rockburst incubation state. Figure 8 It can be seen that the state of the energy accumulation and release stage is not a post-processing judgment, but enters the unified prototype differential attention layer, which constrains the affinity relationship, prototype attribution and multi-source fusion process, thereby enhancing the expressive ability of the stage correspondence.

[0215] See Figure 9 This invention discloses the process by which the input portion performs time alignment, field arrangement, and sequence concatenation on microseismic data, acoustic emission data, stress data, and displacement data. First, four raw inputs are provided, each with a different number of records and field dimensions. Then, time matching is performed according to the sampling time within the underground engineering excavation time window, ensuring that the four sources form a monitoring group at the same sampling time. Next, field alignment is performed according to a fixed source order, ensuring that microseismic, acoustic emission, stress, and displacement data occupy a fixed arrangement position at each sampling time. Finally, time-series concatenation is performed to form a sequence tensor that can be input into a prototype differential attention neural network. Figure 9 As can be seen from this, the present invention ensures that multi-source heterogeneous data have a unified sampling time index and a fixed source relationship before model calculation, thus avoiding confusion in the time position and field position of data from different sources.

[0216] See Figure 10 This study reveals the changes in affinity between precursor monitoring variables and prototypes of six rockburst incubation stages, with and without the energy accumulation and release stage constraint. The horizontal axis represents the sampling time, and the vertical axis represents stages 1 to 6. The color intensity indicates affinity, ranging from 0 to 1. The left-hand heatmap shows the prototype assignment results without the energy accumulation and release stage constraint; the affinity distribution is relatively dispersed, with significant overlap and mixing between different stages, indicating that precursor monitoring variables may have been assigned to incorrect rockburst incubation stage prototypes. The right-hand heatmap shows the results with the energy accumulation and release stage constraint; the affinity exhibits a clearer stage-banded distribution along the sampling time, with stages 1 to 6 showing a more obvious temporal progression. Figure 10 It can be seen that the present invention reconstructs positive and negative affinity channels within the unified prototype differential attention layer of the prototype differential attention neural network: the precursor monitoring variables corresponding to the stage are enhanced to positive affinity, and the precursor monitoring variables that do not correspond to the stage are formed to negative affinity, thereby improving the clarity of prototype attribution and stage distinguishability.

[0217] See Figure 11 This study reveals the correlation between the predicted rockburst incubation stages and the actual stage markings within the 0-24h time window of underground engineering excavation. The first row of the figure shows the actual incubation stages, sequentially including initial accumulation, continuous accumulation, release initiation, release enhancement, instability imminent, and instability release. The second row shows the predicted stages of this invention, along with corresponding confidence levels ranging from 0.60 to 0.98, with higher values ​​indicating higher prediction confidence. The third row shows the stages predicted without this invention's model, revealing numerous stage jumps and unstable stage order in some time periods. In contrast, the predicted stages of this invention have closer stage boundaries to the actual incubation stages, exhibit more continuous overall stage changes, and fewer stage jumps.

[0218] See Figure 12 This invention discloses the process by which the output layer performs response calculations to determine the fused representation of rockburst incubation state. The input on the left is a fusion table, which, after passing through the output layer, enters the type response branch and the stage response branch, respectively. The type response branch forms a type response set and determines the rockburst type determination result based on the maximum response position; the stage response branch forms a stage response set and determines the incubation stage determination result based on the maximum response position. Subsequently, the rockburst type determination result and the incubation stage determination result enter the combination matching module to form a joint result, outputting the rockburst type, incubation stage, and the correspondence between rockburst type and incubation stage. Figure 12It can be seen that the present invention does not perform type identification or stage identification separately, but simultaneously completes the two-category determination based on the fusion representation of rockburst incubation state, and obtains the final fusion result through combination matching.

[0219] Although embodiments of the present invention have been described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present invention, and all such changes and alterations shall not depart from the protection scope of the present invention.

Claims

1. A method for fusing multi-source heterogeneous information on different types of rockburst incubation processes based on deep learning, characterized in that, Includes the following steps: S1. Acquire microseismic data, acoustic emission data, stress data, and displacement data within the time window of underground engineering excavation to form a sequence of monitoring variables for precursors of rockburst incubation; S2. Determine the energy accumulation and release stage state based on the monitoring variable sequence of rockburst precursors. The energy accumulation and release stage state is used to represent the correspondence between elastic energy accumulation and microseismic and acoustic emission energy release. S3. Based on the energy accumulation and release stage state, the unified prototype differential attention layer in the prototype differential attention neural network, which is used to calculate the affinity relationship between precursor monitoring variables and the prototype of the rockburst incubation stage, is reconstructed with positive and negative affinity channels to form energy stage-constrained positive and negative affinity channels. This enables the precursor monitoring variables corresponding to the stage to form a positive affinity relationship with the prototype of the rockburst incubation stage, and enables the precursor monitoring variables that do not correspond to the stage to form a negative affinity relationship with the prototype of the rockburst incubation stage. The prototype differential attention neural network includes: an input mapping layer, a time window encoding layer, a unified prototype differential attention layer, a prototype attribution calculation layer, a multi-source heterogeneous information fusion layer, and an output decision layer; The input mapping layer is used to receive the monitoring variable sequence of rockburst precursors and output the variable representation; The time window encoding layer is used to encode the time-series features of the variable representation and generate time-series features of a unified dimension. The unified prototype differential attention layer has multiple sets of rockburst incubation stage prototypes built in, which can receive the energy accumulation and release stage states to complete the reconstruction of positive and negative affinity channels. The prototype attribution calculation layer completes the stage prototype attribution operation of precursor variables and updates the stage prototype representation based on positive and negative affinity channels; The multi-source heterogeneous information fusion layer adaptively adjusts the multi-source data fusion path according to the prototype representation of each stage, and outputs a fusion representation of the rockburst incubation state. The output determination layer calculates the rockburst type and the determination response of the incubation stage in parallel based on the fusion representation of the rockburst incubation state, and outputs the joint determination result matching the two. S4. Based on the energy stage constraint of positive and negative affinity channels, perform prototype attribution calculation on the monitoring variable sequence of rockburst incubation precursors to form a prototype representation of the rockburst incubation stage; S5. Based on the prototype representation of the rockburst incubation stage, multi-source heterogeneous information fusion is performed to obtain the fused representation of the rockburst incubation state, specifically including steps S51~S54: S51. Receive the prototype representation of the rockburst incubation stage, determine the stage source based on the prototype representation of the rockburst incubation stage to which the precursor monitoring variables belong, and form a prototype representation of the stage source calibration. S52. Based on the stage source calibration prototype representation, calculate the source response of microseismic data, acoustic emission data, stress data, and displacement data in the same rockburst incubation stage prototype, and use the source response to update the fusion basis to form a source response-constrained prototype representation; wherein, the source response is jointly determined by the bottom layer strength belonging to the same rockburst incubation stage prototype and the change in adjacent sampling times: ； Indicates the first Class source in Source response in a prototype of a rockburst incubation stage; Indicates the source type of the current calculation. ; Indicates the source of microseismic data; Indicates the source of acoustic emission data; Indicates the source of stress data; Indicates the source of displacement data; Indicates the prototype index of the rockburst incubation stage; Indicates belonging to the first A set of indexes for sampling times of precursor monitoring variables in the prototype of a rockburst incubation stage; Indicates the sampling time index; Indicates a field index; Indicates the first The set of field indexes corresponding to the class source; Indicates the sampling time The Normalized values ​​for each field; Indicates the sampling time The first adjacent sampling time Normalized values ​​for each field, when At that time, Set as ; This represents the source enumeration variable used to iterate through the four types of sources; Indicates the first The set of field indexes corresponding to the class source; This represents the absolute value operation; This represents the summation operation; S53. Based on the source response constraint prototype representation, compare the source response with the third preset threshold. If the source response is not lower than the third preset threshold, retain the intra-stage fusion path. If the source response is lower than the third preset threshold, suppress the intra-stage fusion path to form an intra-stage fusion representation. S54. Based on the intra-stage fusion representation, multi-source heterogeneous information fusion is performed on the intra-stage fusion representations of different rockburst incubation stage prototypes, and the fusion order is constrained by the stage relationship between the rockburst incubation stage prototypes to obtain the rockburst incubation state fusion representation. S6. Determine the rockburst type and incubation stage based on the fusion representation of rockburst incubation state, and output the fusion results of multi-source heterogeneous information of different types of rockburst incubation processes.

2. The method for fusing multi-source heterogeneous information on different types of rockburst incubation processes based on deep learning as described in claim 1, characterized in that, Step S1 specifically includes: S11. Acquire microseismic data, acoustic emission data, stress data, and displacement data within the underground engineering excavation time window, and determine the time position of the microseismic data, acoustic emission data, stress data, and displacement data within the underground engineering excavation time window; S12. Based on the time location, match the microseismic data, acoustic emission data, stress data and displacement data corresponding to the same sampling time to form a monitoring variable group for the sampling time; S13. Based on the monitoring variable group at the sampling time, set a fixed arrangement position for the source relationship of microseismic data, acoustic emission data, stress data and displacement data, and arrange the monitoring variable group at the sampling time according to the fixed arrangement position of the source relationship to form an aligned monitoring variable group; S14. Connect and align the monitoring variable groups according to the time sequence of the underground engineering excavation time window to form a monitoring variable sequence for precursors of rockburst.

3. The method for fusing multi-source heterogeneous information on different types of rockburst formation processes based on deep learning as described in claim 1, characterized in that, Step S2 specifically includes: S21. Receive the monitoring variable sequence of rockburst precursors, and match the microseismic data, acoustic emission data, stress data and displacement data at the same sampling time according to the time sequence of the underground engineering excavation time window to form a time source variable group; S22. Calculate the stress growth direction, stress growth amplitude, and stress duration between adjacent sampling times based on the stress data in the time source variable group, and map the stress growth direction, stress growth amplitude, and stress duration to the elastic energy accumulation benchmark to form a stage state candidate sequence; S23. Calculate the microseismic energy release change between adjacent sampling times based on the microseismic data in the time source variable group, match the microseismic energy release change with the elastic energy accumulation benchmark in the stage state candidate sequence according to the sampling time, update the stage state candidate sequence according to the matching result, and form a microseismic constraint stage state candidate sequence. S24. Calculate the change in acoustic emission energy release between adjacent sampling times based on the acoustic emission data in the time source variable group. Match the change in acoustic emission energy release with the elastic energy accumulation benchmark in the microseismic constraint stage state candidate sequence according to the sampling time. Update the microseismic constraint stage state candidate sequence according to the matching result to form the microseismic and acoustic emission constraint stage state candidate sequences. S25. Calculate the displacement response change between adjacent sampling times based on the displacement data in the time source variable group, determine the consistency of the displacement response change with the microseismic and acoustic emission constraint stage state candidate sequences in terms of sampling time, update the microseismic and acoustic emission constraint stage state candidate sequences according to the determination results, and form the displacement constraint stage state candidate sequences. S26. The elastic energy accumulation benchmark and the changes in microseismic energy release and acoustic emission energy release at each sampling moment in the candidate sequence of displacement constraint stage states are calibrated to correspond to the changes in elastic energy accumulation benchmark, microseismic energy release, and acoustic emission energy release. The calibration results are compared with the first preset threshold condition. If the first preset threshold condition is met, it is marked as a stage correspondence. If the first preset threshold condition is not met, it is marked as a stage non-correspondence, thus forming an energy accumulation and release correspondence sequence. S27. Determine the energy accumulation and release stage state according to the corresponding sequence of energy accumulation and release, so that the energy accumulation and release stage state represents the correspondence between elastic energy accumulation and micro-vibration and acoustic emission energy release.

4. The method for fusing multi-source heterogeneous information on different types of rockburst incubation processes based on deep learning as described in claim 1, characterized in that, Step S3 specifically includes: S31. At the affinity channel generation position inside the unified prototype differential attention layer of the prototype differential attention neural network, the energy accumulation and release stage state is received, and the initial affinity relationship between the precursor monitoring variable and the prototype of the rockburst incubation stage generated by the unified prototype differential attention layer is obtained, forming the affinity relationship to be reconstructed as the object of positive and negative affinity channel reconstruction. S32. Based on the energy accumulation and release stage state, extract the correspondence between elastic energy accumulation and microseismic and acoustic emission energy release at the affinity channel generation location, and map the correspondence to the precursor monitoring variables and rockburst incubation stage prototypes in the affinity relationship to be reconstructed, forming the initial affinity relationship of stage constraints. S33. Based on the initial affinity relationship of the stage constraint, perform stage correspondence matching between each precursor monitoring variable and each rockburst incubation stage prototype, and calculate the stage correspondence value of the precursor monitoring variable relative to the rockburst incubation stage prototype to form a set of stage correspondence values. The calculation method for the stage-corresponding values ​​of precursor monitoring variables relative to the prototype rockburst incubation stage is as follows: ； in, Indicates the sampling time Precursor monitoring variables relative to the first The corresponding stage values ​​of a prototype rockburst incubation stage; Indicates the sampling time index; Indicates the prototype index of the rockburst incubation stage; Indicates the sampling time Precursor monitoring variables and the first The initial affinity between the prototypes of the rockburst incubation stage; Indicates the sampling time The energy accumulation and release phase state relative to the first Stage constraint results of a prototype rockburst incubation stage; Indicates the sampling time The elastic energy accumulation component; Indicates the sampling time The energy release component of microseismic events; Indicates the sampling time The component of acoustic emission energy released; Indicates the sampling time The displacement response change component; This indicates that the smaller of the two input values ​​is selected. This represents the absolute value operation; S34. Compare the stage corresponding values ​​in the stage corresponding value set with the second preset threshold. If the stage corresponding value is not lower than the second preset threshold, generate a stage corresponding mark. If the stage corresponding value is lower than the second preset threshold, generate a stage non-corresponding mark to form an affinity polarity mark set. S35. Based on the affinity polarity label set, configure positive affinity relationships between precursor monitoring variables with stage-corresponding labels in the initial affinity relationship of stage constraints and the prototype of the rockburst incubation stage, and configure negative affinity relationships between precursor monitoring variables with stage-non-corresponding labels and the prototype of the rockburst incubation stage. Update the affinity direction in the initial affinity relationship of stage constraints with positive and negative affinity relationships to form energy stage constraint affinity relationships. S36. Based on the energy stage constraint affinity relationship, the affinity channels between the precursor monitoring variables and the rockburst incubation stage prototype in the unified prototype differential attention layer are separated into positive and negative channels to form energy stage constraint positive and negative affinity channels. This ensures that the precursor monitoring variables corresponding to the stage form a positive affinity relationship with the rockburst incubation stage prototype, and that the precursor monitoring variables not corresponding to the stage form a negative affinity relationship with the rockburst incubation stage prototype.

5. The method for fusing multi-source heterogeneous information on different types of rockburst incubation processes based on deep learning as described in claim 1, characterized in that, Step S4 specifically includes: S41. Receive the positive and negative affinity channels of the energy stage constraint and the monitoring variable sequence of rockburst incubation precursors. At the affinity reading position of the prototype attribution calculation, pair the precursor monitoring variables in the rockburst incubation precursor monitoring variable sequence with the prototype of the rockburst incubation stage according to the sampling time to form a pairing relationship to be assigned. S42. Based on the energy stage constraint of positive and negative affinity channels, perform channel reading on the positive and negative affinity relationships in the pairing relationships to be assigned, and write the reading results into the assignment determination criteria of the pairing relationships to be assigned, thus forming an energy-constrained pairing relationship to be assigned; S43. Based on the energy-constrained pairing relationship, calculate the positive attribution response for precursor monitoring variables with positive affinity and the prototype of the rockburst incubation stage, and calculate the negative repulsion response for precursor monitoring variables with negative affinity and the prototype of the rockburst incubation stage, thus forming a positive and negative attribution response relationship. Among them, when When it is greater than zero, As a positive attribution response ;when When it equals zero, Set to zero; when When it is greater than zero, As a negative repulsion response ;when When it equals zero, Set to zero; This is the positive affinity channel, indicating the sampling time. Precursor monitoring variables and the first Positive affinity strength between prototypes of the rockburst incubation stage; The negative affinity channel indicates the sampling time. Precursor monitoring variables and the first The negative affinity strength between prototypes in the rockburst incubation stage Indicates the prototype index of the rockburst incubation stage; S44. Based on the positive and negative attribution response relationship, the positive attribution response and negative rejection response of the same precursor monitoring variable relative to the same rockburst incubation stage prototype are calculated by difference to obtain the stage constraint attribution response of each precursor monitoring variable relative to each rockburst incubation stage prototype, forming a stage constraint attribution response set. The calculation method for the stage-constrained attribution response of precursor monitoring variables relative to the prototype of the rockburst incubation stage is as follows: ； in, Indicates the sampling time Precursor monitoring variables relative to the first Stage constraint attribution response of a prototype rockburst incubation stage; Indicates the sampling time Precursor monitoring variables relative to the first Positive attribution response of a prototype of a rockburst incubation stage; Indicates the sampling time Precursor monitoring variables relative to the first The negative repulsion response of a prototype rockburst incubation stage; Indicates the sampling time index; Indicates the prototype index of the rockburst incubation stage; Component indexes representing the time representation and prototype of the rockburst incubation stage; Indicates the sampling time 128-dimensional time representation The Middle One output component; Indicates the first A prototype of the rockburst incubation stage The Middle One prototype component; This indicates that the smaller of the two input values ​​is selected. This indicates that the larger of the two input values ​​is selected. This represents the absolute value operation; This indicates a pre-defined positive stable term; S45. Based on the set of stage constraint attribution responses, perform prototype attribution calculations for each precursor monitoring variable, assign precursor monitoring variables whose stage constraint attribution responses meet the preset attribution conditions to the corresponding rockburst incubation stage prototypes, and exclude precursor monitoring variables whose negative rejection responses meet the preset rejection conditions from the corresponding rockburst incubation stage prototypes, thus forming a stage constraint prototype attribution relationship. S46. Based on the stage-constrained prototype attribution relationship, aggregate the precursor monitoring variables assigned to each rockburst incubation stage prototype, and limit the aggregation range of each rockburst incubation stage prototype with the stage-constrained prototype attribution relationship to form a prototype aggregation representation. S47. Update the prototype of the rockburst incubation stage based on the prototype aggregation representation to form the prototype representation of the rockburst incubation stage.

6. The method for fusing multi-source heterogeneous information on different types of rockburst formation processes based on deep learning as described in claim 1, characterized in that, Step S6 specifically includes: S61. Receive the fused representation of rockburst incubation state, perform judgment response calculation on the fused representation of rockburst incubation state, and form a judgment response set corresponding to the rockburst type and incubation stage respectively; S62. Based on the set of judgment responses, determine the maximum response position of the judgment responses corresponding to the rockburst type to form a rockburst type judgment result, and determine the maximum response position of the judgment responses corresponding to the incubation stage to form an incubation stage judgment result. S63. Combine and match the results of rockburst type determination and incubation stage determination to form a joint determination result corresponding to rockburst type and incubation stage; S64. Output the multi-source heterogeneous information fusion results of different types of rockburst incubation processes based on the joint judgment results.

Images