A Deep Neural Network-Based Intelligent Calculation Method, Equipment, and Medium for Grouting

Abstract

This invention discloses a grouting intelligent calculation method, equipment, and medium based on deep neural networks, relating to the field of neural network technology. The method includes: optimizing the fracture inversion parameter set according to the far-end water-blocking zone and the near-end reinforcement zone of the grouting, and merging it with grouting process parameters and construction constraints to generate a multi-level grouting arrangement input package; inputting the multi-level grouting arrangement input package into a grouting arrangement deep neural network for stage division and grouting process parameter optimization, outputting the stage sequence and pressure-flow trajectory sampling points to generate a stage arrangement list; performing staged grouting according to the stage arrangement list, collecting measured pressure, measured flow rate, and backflow volume, and performing data synchronization and quality control to generate a grouting intelligent calculation record. This invention realizes grouting arrangement under construction constraints, used to dynamically generate pressure-flow trajectory, balancing water-blocking sealing and reinforcement strength.

Description

Technical Field

[0001] This invention relates to the field of neural network technology, and in particular to a grouting intelligent calculation method, device and medium based on deep neural networks. Background Technology

[0002] In the fields of geotechnical engineering and underground structure reinforcement, grouting technology is a crucial means of controlling groundwater seepage and improving the stability of surrounding rock. Its level of intelligence directly impacts construction efficiency and project safety. In recent years, with the rapid development of sensor networks, edge computing, and artificial intelligence technologies, the data acquisition and analysis capabilities of the grouting process have been enhanced, driving a paradigm shift from experience-driven to data-driven approaches. Especially with the rise of deep learning technology, some studies have attempted to use convolutional deep neural networks or recurrent deep neural networks for grouting pressure-flow response modeling, aiming to indirectly identify the state of formation fractures.

[0003] Current mainstream intelligent decision-making systems for grouting have two key limitations: First, in the fracture inversion stage, most models only utilize simplified scalar features such as average pressure or cumulative grouting volume, ignoring the dynamic fracture evolution information contained in the pressure-flow time series response, resulting in limited inversion accuracy; Second, in the grouting scheduling stage, existing methods generally use rule engines or fixed strategies for parameter setting, lacking a differentiable, end-to-end optimization framework, and cannot jointly optimize the division of grouting stages and trajectory generation while meeting construction constraints. Summary of the Invention

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

[0005] Therefore, this invention provides a grouting intelligent calculation method based on deep neural networks to solve the problems of insufficient accuracy in fracture inversion and lack of differentiable optimization in grouting arrangement.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a grouting intelligent calculation method based on a deep neural network, comprising: collecting multi-source operation monitoring data of grouting, performing timestamp alignment, missing data completion, and abnormal jump rejection, and simultaneously extracting pressure and flow response segments according to a fixed sliding window to generate a multi-source aligned response window sequence; based on the multi-source aligned response window sequence, performing convolutional temporal feature extraction and inversion mapping inference in a fracture inversion deep neural network to generate a fracture inversion parameter set; optimizing the fracture inversion parameter set according to the far-end water-blocking zone and the near-end reinforcement zone of grouting, and merging it with grouting process parameters and construction constraints to generate a multi-level grouting arrangement input package; inputting the multi-level grouting arrangement input package into the grouting arrangement deep neural network, performing stage division and grouting process parameter optimization, outputting the stage sequence and pressure and flow trajectory sampling points to generate a stage arrangement list; performing stage grouting according to the stage arrangement list, collecting measured pressure, measured flow, and backflow volume, and performing data synchronization and quality control to generate a grouting intelligent calculation record.

[0007] As a preferred embodiment of the grouting intelligent calculation method based on deep neural networks described in this invention, the steps for generating the multi-source aligned response window sequence are as follows: Collect multi-source operation monitoring data, grouting process parameters and construction constraints, and perform timestamp alignment and time granularity unification. At the same time, bind hole segment identifiers and generate a unified clock reference alignment stream. The unified clock reference alignment stream is used to perform missing measurement completion and abnormal jump rejection, and to maintain the physical consistency between the grouting pump side pressure and the grouting pump side flow rate, thereby generating a quality control alignment stream set; By using a fixed sliding window length and a fixed sliding step size, pressure and flow response segments are extracted from the quality control aligned flow set, and multi-source aligned segments within the same window are extracted simultaneously to generate a multi-source aligned response window sequence.

[0008] As a preferred embodiment of the intelligent grouting calculation method based on deep neural networks described in this invention, the steps for generating the fracture inversion parameter set are as follows: The deep neural network for crack inversion includes an input embedding layer, a one-dimensional convolutional layer, a gated recurrent node layer, and a fully connected mapping layer; The multi-source aligned response window sequence is input into the input embedding layer of the crack inversion deep neural network, and channel rearrangement and window position encoding are performed to generate a unified channel sequence. Based on a unified channel sequence, texture features of abrupt segments and morphological features of steady-state segments are extracted by local convolution along the time axis in a one-dimensional convolutional layer. Based on the texture of the abrupt segment and the morphological features of the steady-state segment, cross-sampling point dependencies are aggregated in the gated loop node layer to generate a temporal representation vector; In the fully connected mapping layer, the temporal representation vector is inverted and projected to generate a fracture inversion parameter set.

[0009] As a preferred embodiment of the intelligent grouting calculation method based on deep neural networks described in this invention, the step of partitioning and optimizing the fracture inversion parameter set includes the following steps: Based on the fracture inversion parameter set, the distal water-blocking zone and the proximal reinforcement zone are divided according to the axial segmentation rules of the grouting hole, and the hole segment parameters are collected and partitioned in combination with the hole segment identifier to generate the hole segment parameter queue. The parameters of the partitioned borehole segments are rearranged and the constraints are associated and organized according to the priority of water plugging and reinforcement, respectively, to generate a partitioned optimization parameter table.

[0010] As a preferred embodiment of the intelligent grouting calculation method based on deep neural networks described in this invention, the steps for generating the multi-level grouting arrangement input package are as follows: The zoning optimization parameter table and grouting process parameters are expanded by unifying the hole segment identification diameter and arranging the process fields to form a merged process parameter table. The process merging parameter table and construction constraints are matched with the hole segment identifier key and the constraint field is concatenated to generate a multi-level grouting arrangement input package.

[0011] As a preferred embodiment of the grouting intelligent calculation method based on deep neural networks described in this invention, the steps for generating the inventory list in the generation stage are as follows: The grouting arrangement deep neural network includes a multi-head self-attention encoding layer, a stage boundary generation layer, a trajectory decoding layer, and a differentiable constraint projection layer; The multi-level grouting arrangement input package is input into the grouting arrangement deep neural network, and query key-value projection and attention convergence are performed in the multi-head self-attention encoding layer to generate a global arrangement representation. In the stage boundary generation layer, the global orchestration representation is deduced into differentiable stages and the stage boundaries are located to generate the stage sequence and stage duration. Based on the phase sequence and phase duration, the phase conditions are expanded and the sampling points are gradually decoded and derived through the trajectory decoding layer. Combined with the construction constraints, the grouting process parameters are optimized to generate pressure and flow trajectory sampling points. In the differentiable constraint projection layer, the pressure and flow trajectory sampling points are continuously projected and the boundaries are clipped to generate a stage orchestration list.

[0012] As a preferred embodiment of the intelligent grouting calculation method based on deep neural networks described in this invention, the steps of performing staged grouting according to the staged arrangement list and collecting measured pressure, measured flow rate, and backflow volume are as follows: The pressure and flow trajectory sampling points are expanded into a control cycle execution sequence according to the phased arrangement list, and the grouting pump control interface is issued. At the same time, the control cycle number is written to form a phased execution instruction. During the control cycle, the measured pressure, measured flow rate, and backflow volume are collected synchronously and aligned according to the control cycle number to generate a three-quantity aligned sampling entry.

[0013] As a preferred embodiment of the grouting intelligent calculation method based on deep neural networks described in this invention, the steps for generating the grouting intelligent calculation record are as follows: The arrival time difference elimination and missing measurement anomaly handling are performed on the three-quantity aligned sampling entries, and the results are collected into a phase execution receipt sequence; Align the phase execution receipt sequence with the phase orchestration list according to the control cycle number, aggregate the actual execution trajectory of the phase, organize it into an execution receipt response window sequence, and encapsulate it into a grouting intelligent calculation record.

[0014] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the grouting intelligent calculation method based on a deep neural network as described in the first aspect of the present invention.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the grouting intelligent calculation method based on a deep neural network as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: by extracting temporal features and mapping inference through a deep neural network for fracture inversion, high-precision fracture parameter inversion is achieved, which is used to accurately characterize the formation state and improve the targeting of grouting and material utilization; by performing stage division and trajectory optimization through a deep neural network for grouting arrangement, grouting arrangement under construction constraints is achieved, which is used to dynamically generate pressure-flow trajectory, taking into account both water plugging and sealing performance and reinforcement strength. Attached Figure Description

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

[0018] Figure 1 This is a flowchart of a grouting intelligent calculation method based on deep neural networks.

[0019] Figure 2 A radar chart comparing comprehensive indicators.

[0020] Figure 3 This is a comparative chart showing the trade-off between water-blocking sealing performance and reinforcement strength.

[0021] Figure 4 This is a comparison chart of pressure and flow trajectory sampling points under construction constraints. Detailed Implementation

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

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

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

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a grouting intelligent calculation method based on a deep neural network, including the following steps: S1: Collect multi-source operation monitoring data for grouting, and perform timestamp alignment, missing data supplementation, and abnormal jump rejection. At the same time, extract pressure and flow response segments according to a fixed sliding window to generate a multi-source aligned response window sequence.

[0026] S1.1: Collect multi-source operation monitoring data of grouting, grouting process parameters and construction constraints, and perform timestamp alignment and time granularity unification. At the same time, bind the hole segment identifier and generate a unified clock reference alignment stream.

[0027] Furthermore, multi-source grouting operation monitoring data, grouting process parameters, and construction constraints are collected, and the data are written into collection timestamps. The timestamp format, time zone caliber, and duplicate / missing information are verified. A time grid scale is established based on a unified time reference, and the sampling points are mapped to time grid indices. Sampling points above the time grid scale are segmented and aggregated according to the time grid index, and sampling points below the time grid scale are interpolated and supplemented according to the time grid index to obtain a unified time granularity sequence. The hole segment identifier and the time grid index are used as alignment keys to associate the multi-source grouting operation monitoring data, grouting process parameters, and construction constraints, and encapsulated into alignment entries to generate a unified clock reference alignment stream.

[0028] It should be noted that the time raster scale (e.g., 1-second granularity) aligns the sampling time according to the time raster scale, ensuring that the data points within each second are correctly mapped to the corresponding time raster index.

[0029] The multi-source grouting operation monitoring data includes the grouting pump side pressure time series, grouting pump side flow time series, and backflow time series, and is accompanied by acquisition timestamps, hole segment identifiers, and control cycle numbers for alignment and organization.

[0030] Grouting process parameters include the proportion and dosage of grouting materials organized according to the hole segment identification, the target pressure and target flow rate setting, the stage construction arrangement settings, and the set of fields issued by the grouting pump control interface.

[0031] Construction constraints include a set of constraint fields such as pressure and flow boundaries, stage duration limits, equipment capacity boundaries, and safety boundaries, which are used for constraint matching and boundary trimming in the stage division and trajectory output process.

[0032] S1.2: Perform missing measurement completion and abnormal jump rejection on the unified clock reference alignment stream, and maintain the physical consistency between the grouting pump side pressure and the grouting pump side flow rate to generate a quality control alignment stream set.

[0033] Furthermore, missing data is filled into the unified clock reference alignment stream by using linear interpolation to fill in missing values ​​in the time series, ensuring that there is corresponding sampling data at each time point; abnormal jumps are removed by identifying and deleting abrupt changes in the time series (such as values ​​with a change of more than 5%) to ensure data stability; physical consistency checks are performed on the grouting pump side pressure and grouting pump side flow data to ensure that the temporal correlation and physical relationship between the two data are not disrupted; and the processed data are used to generate a quality control alignment stream set.

[0034] It should be noted that physical consistency is determined by synchronously verifying the direction and magnitude of change of the grouting pump side pressure and the grouting pump side flow rate at the same time scale. When the grouting pump side pressure rises while the grouting pump side flow rate continues to decrease, or when the grouting pump side flow rate is zero but the grouting pump side pressure continues to rise and abnormal combinations occur continuously, it is determined to be physical inconsistency and triggers the handling.

[0035] S1.3: Extract pressure and flow response segments from the quality control aligned flow set according to a fixed sliding window length and a fixed sliding step size, and simultaneously extract multi-source aligned segments in the same window to generate a multi-source aligned response window sequence.

[0036] Furthermore, based on the fixed sliding window length and fixed sliding step size, a list of window sequence numbers and window time boundaries is established. According to the window time boundary list, the time index range covered by each window is located in the quality control alignment flow set. The grouting pump side pressure and grouting pump side flow rate within the time index range are extracted and encapsulated as pressure-flow response segments. Multi-source alignment entries in the quality control alignment flow set are synchronously extracted using the same time index range and encapsulated as multi-source alignment segments within the same window. The pressure-flow response segments and multi-source alignment segments within the same window are merged according to the window sequence number, and the hole segment identifier and window time boundary are registered to generate a multi-source alignment response window sequence.

[0037] It should be noted that the fixed sliding window length (example value: 60s) and fixed sliding step size (example value: 10s) are determined by the grouting condition decision time scale and selected according to the stable observation duration and control update frequency of the pressure and flow rate changes during the grouting process, so that each pressure and flow rate response segment covers a identifiable condition segment and that adjacent pressure and flow rate response segments advance continuously.

[0038] The multi-source aligned response window sequence slices the grouting process according to a unified time window and organizes the pressure and flow response segments and the multi-source aligned segments in the same window into a continuous input sequence of a deep neural network under the same window number caliber. This sequence is used to support the inference generation of the fracture inversion parameter set and the stage arrangement list.

[0039] S2: Based on the multi-source aligned response window sequence, convolutional temporal feature extraction and inversion mapping inference are performed in the deep neural network for crack inversion to generate a crack inversion parameter set.

[0040] S2.1: The deep neural network for crack inversion includes an input embedding layer, a one-dimensional convolutional layer, a gated recurrent node layer, and a fully connected mapping layer.

[0041] Furthermore, the deep neural network for crack inversion consists of an input embedding layer, a one-dimensional convolutional layer, a gated recurrent node layer, and a fully connected mapping layer cascaded in a sequential manner. The input embedding layer receives the multi-source aligned response window sequence and organizes the multi-source channels into a unified channel sequence. The one-dimensional convolutional layer extracts local temporal morphological features along the time axis from the unified channel sequence and outputs a convolutional feature sequence. The gated recurrent node layer aggregates the cross-sampling point dependencies of the convolutional feature sequence and outputs a temporal representation vector. The fully connected mapping layer maps the temporal representation vector to the inversion output space to form the output representation of the crack inversion parameter set.

[0042] It should be noted that the training of the deep neural network for fracture inversion is based on the supervision sample pairs of historical multi-source aligned response window sequences and corresponding fracture inversion parameter sets. The multi-source aligned response window sequences are rearranged and the window positions are encoded in the input embedding layer to obtain a unified channel sequence, which is then pushed sequentially to a one-dimensional convolutional layer, a gated recurrent node layer, and a fully connected mapping layer to obtain the fracture inversion parameter set expression of the network output. Error calculation is performed on the fracture inversion parameter set in the network output and the supervision sample pairs, and time continuity constraints and grouting pump side pressure-grouting pump side flow consistency constraints are superimposed to form the training objective. The parameters of the input embedding layer, one-dimensional convolutional layer, gated recurrent node layer, and fully connected mapping layer are updated through backpropagation, and the iteration is performed in batches according to the window number until the validation set error converges. At the same time, the output error distribution after training is segmented and statistically analyzed to solidify the inversion confidence label caliber for use in the subsequent inference stage output fracture inversion parameter set.

[0043] S2.2: Input the multi-source aligned response window sequence into the input embedding layer of the crack inversion deep neural network, perform channel rearrangement and window position encoding, and generate a unified channel sequence.

[0044] Furthermore, the multi-source aligned response window sequence is traversed by window number, and the pressure-flow response segment corresponding to each window and the multi-source aligned segment in the same window are extracted. According to the channel index table of the fracture inversion deep neural network, the pressure-flow response segment and the multi-source aligned segment in the same window are rearranged into channels. The multi-source channels are spliced ​​into a unified time axis channel matrix and the orifice segment identifier and window time boundary are registered. The unified time axis channel matrix is ​​used to generate window position codes according to the time index within the window and written into the position mark vector. The window position codes correspond one-to-one with the time index and are aligned with the channel dimension. The channel rearrangement result and the window position code are merged and encapsulated with the same tensor caliber to generate a unified channel sequence.

[0045] It should be noted that window position encoding involves attaching a set of position marker vectors aligned with the channel dimension to each sampling point within the sliding window, based on the time index. This is used to distinguish the order and relative interval of different time positions within the same sliding window.

[0046] The channel index table is defined as an index list that establishes a fixed mapping relationship between the pressure and flow response segments in the multi-source aligned response window sequence and the multi-source aligned segments in the same window according to "channel name - channel number - sampling time index". The channel index table is fixed according to the necessary order of data fields in the grouting condition decision caliber and the consistent sampling time index length.

[0047] S2.3: Based on a unified channel sequence, local convolution along the time axis is used in a one-dimensional convolutional layer to extract texture features of abrupt transition segments and morphological features of steady-state segments.

[0048] Furthermore, the unified channel sequence is input into a one-dimensional convolutional layer according to the window number, while keeping the channel dimension and time index dimension unchanged. Based on the kernel length and stride of the local convolution on the time axis, neighboring time slices are slid along the time axis and weighted converged with the kernel weights to calculate the convolutional response sequence value. The local difference amplitude and local energy concentration of the convolutional response sequence are statistically analyzed according to the time index and written into abrupt response markers to solidify the texture of abrupt segments. The local stationarity and morphological consistency of the convolutional response sequence are statistically analyzed according to the time index and written into steady-state response markers to solidify the morphological features of steady-state segments. The abrupt response markers, steady-state response markers, and convolutional response sequences are encapsulated into a one-dimensional convolutional layer output feature sequence according to the window number.

[0049] It should be noted that abrupt segment texture refers to a temporal fine-grained change pattern in the output feature sequence of a one-dimensional convolutional layer, where the local difference amplitude is concentrated and the local energy rapidly accumulates within a short time index range. It is used to characterize the abrupt change segment features of pressure and flow response segments.

[0050] Steady-state segment morphological features refer to temporal contour features in the output feature sequence of a one-dimensional convolutional layer that have high local stability and stable morphological consistency within the continuous time index range. They are used to characterize the stable maintenance segment features of pressure and flow response segments.

[0051] The mutation response marker is an indication information that marks the sampling positions with large local difference amplitude and sudden increase in local energy concentration on the time index of the convolutional response sequence. It is used to locate the texture location range of the mutation segment in the pressure and flow response segment.

[0052] The formula for calculating the numerical value of the convolutional response sequence is: ; in, This indicates that the convolutional response sequence is in window number 1. The output channel number is The output position index is The output value at that location, The unified channel sequence is indicated by the window number. Channel number is Input values ​​at the location, This indicates the number of channels contained in a unified channel sequence. This indicates the number of sampling points covered by the convolution kernel on the time axis. Indicates output channel In the input channel Upper and inner core positions The corresponding weighting coefficients, Indicates output channel The corresponding bias value, Indicates the output position index The propulsion interval mapped on the input time axis. This indicates the alignment compensation offset of the input time index at the boundary. Indicates the window number. Indicates the output position index. Indicates the position index within the convolution kernel. Indicates the input channel number. Indicates the output channel number.

[0053] It should be noted that the weight coefficients are obtained by performing forward computation on the sample sequence and iteratively updating it based on the output error during the training phase of the convolutional network, so that the corresponding convolutional kernel weights are stable at the optimal values ​​for the current task when the loss converges.

[0054] The alignment compensation offset is calculated based on the "input time index anchor point corresponding to the output position index" after given the convolution kernel coverage length, stride and boundary padding method. It is used to compensate the alignment position of the convolution kernel at the boundary to the desired time reference frame.

[0055] S2.4: Based on the texture of the abrupt segment and the morphological features of the steady-state segment, aggregate cross-sampling point dependencies in the gated loop node layer to generate a temporal representation vector.

[0056] Furthermore, the output feature sequence of the one-dimensional convolutional layer is input into the gated recurrent node layer according to the window number and expanded into a sample-by-sample input sequence according to the time index. The gated recurrent node layer maintains the hidden state for each time index and performs update, preservation and reset of the hidden state according to the gate weight. The gate weight is obtained by the sample-by-sample input sequence and the hidden state of the previous time index. The mutation response marker corresponding to the texture of the mutation segment is used to improve the update intensity of the hidden state and increase the memory retention of the mutation neighborhood. The steady-state response marker corresponding to the morphological features of the steady-state segment is used to suppress short-term fluctuations and stabilize the continuous accumulation of the hidden state. The gated recurrent node layer recursively converges the cross-sample-point dependencies along the time index to obtain the terminal hidden state and the full sequence converged state. The terminal hidden state and the full sequence converged state are concatenated or weighted to obtain a fixed-dimensional output and encapsulated according to the window number to generate a temporal representation vector.

[0057] S2.5: In the fully connected mapping layer, the temporal representation vector is inverted and projected to generate a fracture inversion parameter set.

[0058] Furthermore, the time-series representation vector is input into the fully connected mapping layer according to the window number and associated with the borehole segment identifier and window time boundary to establish an output index. The fully connected mapping layer configures the weight matrix and bias vector according to the inversion output dimension and performs linear projection on the time-series representation vector to obtain the inversion output vector. The numerical range is clipped and the unit diameter is normalized on the inversion output vector to meet the constraints of the grouting working condition. The inversion output vector is unpacked into fracture inversion parameter entries according to the predetermined parameter order and the window number, borehole segment identifier and window time boundary are registered. All fracture inversion parameter entries are aggregated and encapsulated into a fracture inversion parameter set.

[0059] It should be noted that the weight matrix is ​​used to linearly converge the components of the time series representation vector according to the required combination relationship of the inversion output dimension to obtain the inversion output vector. The shape of the weight matrix is ​​determined by the vector length of the time series representation vector and the vector length of the inversion output vector and is updated and fixed during the training process.

[0060] The bias vector is used to provide a learnable baseline translation on each inversion output component of the inversion output vector to compensate for the overall offset and improve the mapping expressiveness. The length of the bias vector is consistent with the vector length of the inversion output vector and is updated and fixed during training.

[0061] S3: Based on the far-end water-blocking zone and the near-end reinforcement zone of the grouting, the fracture inversion parameter set is partitioned and optimized, and then merged with the grouting process parameters and construction constraints to generate a multi-level grouting arrangement input package; S3.1: Based on the fracture inversion parameter set, divide the far-end water-blocking zone and the near-end reinforcement zone according to the axial segmentation rules of the grouting hole, and combine the hole segment identifier to collect and partition the zone, generating a hole segment parameter queue.

[0062] Furthermore, the fracture inversion parameter set is merged according to the borehole segment identifier, and the fracture inversion parameter entry sequence corresponding to each borehole segment identifier is extracted. According to the axial segmentation rule of grouting holes, the borehole segment identifier is mapped to the axial position interval of the borehole and the interval boundary mark is registered. According to the interval boundary mark, the fracture inversion parameter entry sequence is divided into the far-end water-blocking zone entry group and the near-end reinforcement zone entry group and written into the partition mark. The far-end water-blocking zone entry group and the near-end reinforcement zone entry group are sorted according to the window time boundary, and the gap interval is filled and sorted by the adjacent entry to form the partition continuous entry sequence. The borehole segment identifier, partition mark and partition continuous entry sequence are encapsulated into queue entries and aggregated into the borehole segment parameter queue.

[0063] It should be noted that the axial segmentation rule for grouting holes is used to divide the axial position interval of the same hole segment into a far-end water-blocking zone and a near-end reinforcement zone to support the optimized arrangement of the zones. The axial segmentation rule for grouting holes is based on the hole depth mileage benchmark corresponding to the hole segment identifier and combined with the main seepage control interval of the far-end water-blocking zone and the main rock support control interval of the near-end reinforcement zone to determine the segment boundary and solidify it as the hole depth interval diameter (for example, with the hole opening as 0m and the hole bottom as 50m, 0m-30m is defined as the near-end reinforcement zone and 30m-50m is defined as the far-end water-blocking zone).

[0064] S3.2: Perform parameter rearrangement and constraint association organization on the parameter queues of the partitioned borehole segments for priority of water plugging and priority of reinforcement respectively, and generate partitioned optimization parameter tables.

[0065] Furthermore, the borehole segment parameter queue is split into a far-end water-blocking zone queue and a near-end reinforcement zone queue according to the partition marker. In the far-end water-blocking zone queue, the fracture inversion parameter entries are sorted according to the seepage sensitivity and abrupt response intensity, and a water-blocking priority order marker is generated. Adjacent entries are merged into continuous water-blocking segments and written into the segment boundary and representative entries within the segment. In the near-end reinforcement zone queue, the fracture inversion parameter entries are sorted according to the steady-state response intensity and borehole axial continuity, and a reinforcement priority order marker is generated. Adjacent entries are merged into continuous reinforcement segments and written into the segment boundary and representative entries within the segment. The water-blocking priority order marker and the reinforcement priority order marker are associated with the borehole segment identifier, partition marker, and borehole axial position interval, respectively. Priority pruning is performed on the order conflicting entries. The sorted water-blocking segments and reinforcement segments are aggregated into a partitioned optimization parameter table according to the borehole segment identifier.

[0066] Figure 2 To compare the overall performance of different methods in terms of comprehensive beneficial effects, the indicators include compliance rate, trajectory smoothness, constraint violation rate, compliance water-blocking sealing performance, compliance reinforcement strength, and compliance material utilization rate. The radar chart presents the relative levels of each indicator using a unified dimensional normalization method. The curves correspond to the baseline method, the differentiable grouting orchestration deep neural network (unconstrained projection), and the differentiable grouting orchestration deep neural network (differentiable constrained projection layer), respectively. From the curve envelope, it can be observed that the differentiable constrained projection layer maintains a higher level in compliance rate and trajectory smoothness, while corresponding to a lower value in the constraint violation rate dimension. This reflects that the pressure and flow trajectory sampling points output by the staged orchestration list are more likely to meet boundary requirements and maintain continuity within the stage under construction constraints. The simultaneous improvement in compliance water-blocking sealing performance, compliance reinforcement strength, and compliance material utilization rate indicators supports the comprehensive effect of "balancing water-blocking sealing performance and reinforcement strength, while improving material utilization rate."

[0067] S3.3: The zoning optimization parameter table and grouting process parameters are expanded by unifying the hole segment identification diameter and arranging the process fields to form a merged process parameter table.

[0068] Furthermore, the partition optimization parameter table is traversed by hole segment identifier to extract partition markers, hole axial position intervals, and partition boundary entries. A hole segment identifier mapping table is established for the grouting process parameters, and the hole segment identifier format, numbering granularity, and default values ​​are checked. The grouting process parameters are converted by the hole segment identifier mapping table to the same hole segment identifier caliber as the partition optimization parameter table. The grouting process parameters are expanded into a list of process fields into a sequence of fields that can be spliced ​​and written with field order markers. The splicable field sequences under the same hole segment identifier are deduplicated, merged, and conflicting fields are trimmed. The partition boundary entries in the partition optimization parameter table are aligned and spliced ​​with the splicable field sequences in the grouting process parameters according to the hole segment identifier, and the partition markers and hole axial position intervals are appended to the splicing result to form a process merging parameter table.

[0069] S3.4: Perform hole segment identifier key matching and constraint field splicing on the process merge parameter table and construction constraints to generate a multi-level grouting arrangement input package.

[0070] Furthermore, a borehole segment identifier index is established for the construction constraints, and a constraint field splicing list is compiled. The construction constraints are grouped into a constraint entry sequence according to the borehole segment identifier index, and a default flag is registered for the default constraints (e.g., using a general constraint caliber). The process merged parameter table is traversed according to the borehole segment identifier and the partition flag, borehole axial position interval, partition boundary entries, and splicable field sequences are extracted. Borehole segment identifier key matching is performed to locate the corresponding constraint entry sequence. For borehole segment identifiers that fail to match, a matching anomaly flag is registered, and constraint placeholders are filled in according to the constraint field splicing list. The constraint fields in the constraint entry sequence are expanded into a constraint field sequence according to the constraint field splicing list and written into the field order flag. The unit caliber is unified and the boundary value is trimmed for the constraint field sequence. The constraint field sequence and the process merged parameter table entries are spliced ​​according to the borehole segment identifier and the partition flag and borehole axial position interval are encapsulated to generate a multi-level grouting arrangement input package.

[0071] S4: Input the multi-level grouting arrangement input package into the grouting arrangement deep neural network, perform stage division and grouting process parameter optimization, output the stage sequence and pressure and flow trajectory sampling points, and generate a stage arrangement list; S4.1: The grouting arrangement deep neural network includes a multi-head self-attention encoding layer, a stage boundary generation layer, a trajectory decoding layer, and a differentiable constraint projection layer.

[0072] Furthermore, the grouting arrangement deep neural network is composed of a multi-head self-attention encoding layer, a stage boundary generation layer, a trajectory decoding layer, and a differentiable constraint projection layer cascaded in a sequential manner. The multi-head self-attention encoding layer receives multi-level grouting arrangement input packets and aggregates hole segment identifiers, partition markers, and constraint field sequences to form a global arrangement representation. The stage boundary generation layer provides the stage sequence and stage duration based on the global arrangement representation to determine the stage division criteria. The trajectory decoding layer generates pressure and flow trajectory sampling points based on the stage sequence and stage duration to express the stage execution trajectory. The differentiable constraint projection layer performs continuous projection and boundary trimming on the pressure and flow trajectory sampling points to meet the construction constraints and outputs a constraint-consistent expression of the stage arrangement list.

[0073] It should be noted that the training of the grouting arrangement deep neural network is based on the supervision sample pairs of historical multi-level grouting arrangement input packages and corresponding stage arrangement lists. The multi-level grouting arrangement input packages are sequentially pushed to the multi-head self-attention encoding layer, stage boundary generation layer, trajectory decoding layer, and differentiable constraint projection layer to obtain the stage arrangement list output by the network. Error calculation is performed on the stage order, stage duration, and pressure-flow trajectory sampling points output by the network and the stage arrangement list in the supervision sample pairs, and stage boundary smoothing constraint terms and trajectory continuity constraint terms are added. At the same time, violation penalty terms are superimposed on the boundary consistency between the output of the differentiable constraint projection layer and the construction constraint conditions to form the training objective. The parameters of the multi-head self-attention encoding layer, stage boundary generation layer, trajectory decoding layer, and differentiable constraint projection layer are updated through backpropagation and iterated in batches according to the hole segment identifier until the validation set error converges. After training, the error distribution under different partition labels is segmented and statistically analyzed to solidify the stage boundary stability caliber and trajectory executability caliber for subsequent inference stage output of the stage arrangement list.

[0074] S4.2: Input the multi-level grouting arrangement input package into the grouting arrangement deep neural network, and perform query key-value projection and attention convergence in the multi-head self-attention encoding layer to generate a global arrangement representation.

[0075] Furthermore, the multi-level grouting arrangement input package is expanded into an input item sequence according to the hole segment identifier and the partition marker, hole axial position interval and constraint field sequence are registered. Vectorization embedding is performed on the input item sequence to obtain the arrangement input sequence. In the multi-head self-attention encoding layer, query projection, key projection and value projection are performed on the arrangement input sequence to form query vector sequence, key vector sequence and value vector sequence respectively. The similarity between the query vector sequence and the key vector sequence is obtained according to the multi-head partition to obtain the attention weight matrix. The attention weight matrix is ​​normalized and the value vector sequence is weighted and converged to obtain the multi-head attention output sequence. The multi-head attention output sequence is head splicing and linear integration is performed to obtain the encoded output sequence. The encoded output sequence is pooled and converged according to the hole segment identifier and partition marker to form a global convergence vector. The encoded output sequence, global convergence vector and index marker are encapsulated to generate a global arrangement representation.

[0076] It should be noted that the global orchestration representation is a unified orchestration vector expression obtained by the multi-head self-attention coding layer by converging the sequence of hole segment entries in the multi-level grouting orchestration input package under the cross-hole segment dependency and constraint association caliber. It is used to provide the stage boundary generation layer with global condition information required for stage order, stage duration and trajectory decoding.

[0077] Pooling convergence compresses multiple sequence vectors corresponding to the same hole segment identifier and the same partition label in the encoded output sequence into a single global converged vector by performing max pooling or mean pooling along the sequence dimension. This reduces the sequence length while maintaining the main orchestration features and forms a converged representation of the global orchestration.

[0078] Figure 3 This diagram depicts the trade-off distribution of multiple sample groups regarding water-stopping sealing performance and reinforcement strength. The horizontal axis represents the water-stopping sealing performance, and the vertical axis represents the reinforcement strength. Different colored scatter points correspond to the differentiable grouting orchestration deep neural network (unconstrained projection), the differentiable grouting orchestration deep neural network (differentiable constrained projection layer), and the baseline method, respectively. Each scatter point corresponds to a stage entry or summary sample under a borehole segment identifier, used to map the stage orchestration list and the recordable results from the intelligent grouting calculation record to a two-dimensional effect space. The concentration of the scatter point distribution reflects the output stability; the clustering trend towards the upper right reflects the degree of achievement in "simultaneously considering water-stopping sealing performance and reinforcement strength." The decrease in vertical dispersion within the same water-stopping sealing performance range reflects the suppression of strength result fluctuations by stage division and trajectory optimization. The scatter point distribution corresponding to the differentiable constrained projection layer is closer to the feasible range and maintains a better balance, used to demonstrate the supporting role of adaptive grouting orchestration in the trade-off results under construction constraints.

[0079] S4.3: In the stage boundary generation layer, the global orchestration representation is deduced into differentiable stages and the stage boundaries are located, generating the stage order and stage duration.

[0080] Furthermore, the global orchestration representation is input to the stage boundary generation layer and index-aligned with the hole segment identifier and partition label. The stage boundary generation layer performs stage score sequence prediction on the global convergence vector to obtain continuous stage probability curves, and simultaneously predicts the boundary activation sequence on the encoded output sequence to obtain the boundary activation curve. Differentiable boundary selection is performed on the boundary activation curve to generate boundary position weights and the boundary position weights are merged into the stage boundary positioning result. Differentiable sequential decoding is performed on the continuous stage probability curve to obtain the stage sequence label sequence. The stage time span is obtained from the stage boundary positioning result according to the adjacent boundary position weights and converted into stage duration. The stage sequence label sequence and stage duration are encapsulated according to the hole segment identifier and written into the boundary confidence label to generate the stage sequence and stage duration.

[0081] S4.4: Based on the stage sequence and stage duration, the stage conditions are expanded and the sampling points are gradually decoded and derived through the trajectory decoding layer. Combined with the construction constraints, the grouting process parameters are optimized to generate pressure and flow trajectory sampling points.

[0082] Furthermore, the stage sequence and stage duration are input into the trajectory decoding layer according to the hole segment identifier and expanded into a stage condition sequence. The stage condition sequence includes the stage number, stage duration, and partition mark, and a constraint index is established with the construction constraints. The trajectory decoding layer generates a time index scale within the stage according to the stage condition sequence and outputs a candidate sequence of sampling points recursively according to the time index. The candidate sequence of sampling points gives a pressure sampling point and a flow sampling point for each time index and writes the sampling point number. Constraint violation mark and constraint margin mark are added to the candidate sequence of sampling points according to the construction constraints. For the candidate entries of sampling points with constraint violation mark, optimization and update are performed and adjacent candidate entries of sampling points are adjusted synchronously to maintain trajectory continuity. The optimized candidate sequence of sampling points is merged according to the stage number and the start and end time boundaries of the stage are registered to generate pressure and flow trajectory sampling points.

[0083] It should be noted that the constraint violation marker is an indication message in the sampling point candidate sequence that marks the pressure sampling point or flow sampling point that exceeds the allowable range of the construction constraint conditions. It is used to trigger the optimization of grouting process parameters and the synchronous adjustment of adjacent sampling points.

[0084] Grouting process parameters include stage target pressure setting values ​​(e.g., 2.5MPa, the upper limit of the maximum allowable pressure boundary and equipment capacity boundary given by construction constraints) and stage target flow rate setting values ​​(e.g., 30L / min, the flow rate range that can be stably output according to equipment capacity boundaries), which are used to constrain the value range of pressure sampling points and flow rate sampling points in each stage of the candidate sequence of sampling points.

[0085] Figure 4This diagram illustrates the differences in pressure and flow trajectory sampling points across different methods and the effects of constrained projection in the staged orchestration list. The horizontal axis represents the sampling point number, and the vertical axis represents pressure. The diagram shows three pressure and flow trajectories: the baseline method (sampling point pressure), the differentiable grouting orchestration deep neural network (unconstrained projection), and the differentiable grouting orchestration deep neural network (differentiable constrained projection layer). A dashed line overlaying the construction constraint condition (pressure upper bound) serves as a boundary reference. In the overview diagram, red dashed rectangles mark the intervals of lower-order sampling points and connect them to the enlarged view below. The enlarged view compares the pressure entries within the same window between the unconstrained projection and the differentiable constrained projection layer, highlighting the points of maximum difference and their corresponding values. The pressure entries output by the differentiable constrained projection layer better fit the pressure upper bound and maintain continuous variation, demonstrating the supporting role of continuous projection updates and boundary trimming in ensuring trajectory feasibility and quality within a stage.

[0086] S4.5: In the differentiable constraint projection layer, perform continuous projection and boundary trimming on the pressure and flow trajectory sampling points to generate a stage orchestration list.

[0087] Furthermore, the pressure and flow trajectory sampling points are input into the differentiable constraint projection layer and indexed according to the hole segment identifier and stage number. The difference between adjacent sampling points is obtained according to the sampling point number of the pressure and flow trajectory sampling points, and a continuous constraint quantity sequence is generated. Based on the continuous constraint quantity sequence, continuous projection update is performed on sampling points with excessive jumps, and adjacent sampling points are synchronously written back to maintain the smoothness of the trajectory within the stage. A boundary constraint quantity sequence is generated for the pressure and flow trajectory sampling points according to the construction constraint conditions. Based on the boundary constraint quantity sequence, boundary clipping is performed on sampling points that exceed the upper limit of pressure, are below the lower limit of pressure, exceed the upper limit of flow, and are below the lower limit of flow, and clipping marks are written. The continuous projection update results and boundary clipping results are merged according to the stage number and the stage sequence, stage duration, and stage start and end time boundaries are registered, encapsulated into a stage entry sequence, and aggregated to generate a stage arrangement list.

[0088] S5: Based on the phased arrangement list, perform phased grouting, collect measured pressure, measured flow rate and backflow volume, and perform data synchronization and quality control to generate intelligent grouting calculation records; S5.1: The pressure and flow trajectory sampling points are expanded into a control cycle execution sequence according to the phased arrangement list, and the grouting pump control interface is issued. At the same time, the control cycle number is written to form a phase execution instruction.

[0089] Furthermore, the stage arrangement list is traversed by borehole segment identifier to extract the stage sequence, stage duration, stage start and end time boundaries, and pressure and flow trajectory sampling points. Based on the stage duration and control cycle duration, a control cycle scale is generated, and the pressure and flow trajectory sampling points are mapped to the control cycle scale to form a control cycle execution sequence. The control cycle execution sequence is written with a control cycle number according to the control cycle scale, and the pressure and flow sampling points corresponding to each control cycle number are encapsulated into control instruction entries. The control instruction entries are appended with borehole segment identifier, stage number, and effective time marker. The control instruction entries are sorted by control cycle number and packaged into stage execution instructions, which are sent to the grouting pump control interface and the sent status marker is written back to form the stage execution instructions.

[0090] S5.2: During the control cycle, the measured pressure, measured flow rate and backflow volume are collected synchronously and aligned according to the control cycle number to generate a sampling entry with the three quantities aligned.

[0091] Furthermore, measured pressure, measured flow rate, and backflow volume are synchronously collected within each control cycle. Each sampling point is time-synchronized and bound to the control cycle number. Measured pressure, measured flow rate, and backflow volume are sampled one by one within each control cycle, ensuring that the sampling time is consistent with the control cycle and that each data point has a corresponding time within the control cycle. The collected data are organized according to the control cycle number, aligning the pressure, flow rate, and backflow volume data points by time. If missing data exists, interpolation methods are used to fill in the gaps, and abnormal data caused by abnormal fluctuations are removed. The pressure, flow rate, and backflow volume values ​​of each sampling point are encapsulated into a three-quantity aligned sampling entry.

[0092] S5.3: Perform arrival time difference elimination and missing measurement anomaly handling on the three-quantity aligned sampling entries, and collect them into a phase execution receipt sequence.

[0093] Furthermore, arrival time difference elimination is performed on the three-quantity aligned sampling entries to obtain the difference between the actual arrival time and the predetermined time for each sampling point, and anomalies in the difference are marked. Anomalies are removed, and missing data points are filled in using linear interpolation to ensure that all three-quantity sampling entries within each control cycle are valid data. The valid sampling entries after eliminating time differences and missing anomalies are sorted by control cycle number and collected into a phase execution receipt sequence. Each sampling entry contains corrected pressure, flow, and backflow data, as well as the corresponding control cycle number, timestamp, and marker, forming a complete phase execution receipt sequence.

[0094] It should be noted that the predetermined time is defined as the expected arrival time of the control cycle, derived by combining the control cycle duration and the control cycle sequence number with the control cycle zero-point time marker registered when the stage execution instruction is issued.

[0095] S5.4: Align the stage execution receipt sequence with the stage orchestration list according to the control cycle number, aggregate the actual execution trajectory of the stage, organize it into an execution receipt response window sequence, and encapsulate it into a grouting intelligent calculation record.

[0096] It should be noted that the phase execution receipt sequence and the phase orchestration list are aligned according to the control cycle number to ensure that each receipt item is consistent with the corresponding phase orchestration instruction data item. Successfully matched items are sorted according to the control cycle number. Unmatched items are filled by interpolation or deleted to ensure the integrity of execution data within each control cycle. The aligned phase execution receipt items are aggregated, and the actual execution data of pressure, flow, and return volume are extracted and compared with the expected phase execution trajectory. The compared actual phase execution trajectory is organized into an execution receipt response window sequence according to the control cycle number to ensure time sequence consistency and remove redundant data. The execution receipt response window sequence is summarized and packaged into a grouting intelligent calculation record according to the control cycle.

[0097] This embodiment also provides a computer device applicable to the grouting intelligent calculation method based on deep neural networks, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the grouting intelligent calculation method based on deep neural networks proposed in the above embodiment.

[0098] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0099] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the grouting intelligent computing method based on a deep neural network as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0100] In summary, this invention achieves high-precision fracture parameter inversion through temporal feature extraction and mapping inference using a fracture inversion deep neural network, which is used to accurately characterize the formation state and improve the targeting of grouting and material utilization. It also achieves grouting arrangement under construction constraints through stage division and trajectory optimization using a grouting arrangement deep neural network, which is used to dynamically generate pressure-flow trajectory, taking into account both water plugging and reinforcement strength.

[0101] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A grouting intelligent calculation method based on deep neural networks, characterized in that: include, Collect multi-source operation monitoring data for grouting, and perform timestamp alignment, missing data supplementation, and abnormal jump rejection. At the same time, extract pressure and flow response segments according to a fixed sliding window to generate a multi-source aligned response window sequence. Based on the multi-source aligned response window sequence, convolutional temporal feature extraction and inversion mapping inference are performed in the deep neural network for crack inversion to generate a crack inversion parameter set. Based on the far-end water-blocking zone and the near-end reinforcement zone of the grouting, the fracture inversion parameter set is partitioned and optimized, and then merged with the grouting process parameters and construction constraints to generate a multi-level grouting arrangement input package; The multi-level grouting arrangement input package is input into the grouting arrangement deep neural network to perform stage division and grouting process parameter optimization, outputting the stage sequence and pressure and flow trajectory sampling points to generate a stage arrangement list. According to the phased grouting list, phased grouting is carried out, and the measured pressure, measured flow rate and backflow volume are collected. Data synchronization and quality control are performed, and intelligent calculation records for grouting are generated.

2. The intelligent grouting calculation method based on deep neural networks as described in claim 1, characterized in that: The steps for generating the multi-source aligned response window sequence are as follows: Collect multi-source operation monitoring data, grouting process parameters and construction constraints, and perform timestamp alignment and time granularity unification. At the same time, bind hole segment identifiers and generate a unified clock reference alignment stream. The unified clock reference alignment stream is used to perform missing measurement completion and abnormal jump rejection, and to maintain the physical consistency between the grouting pump side pressure and the grouting pump side flow rate, thereby generating a quality control alignment stream set; By using a fixed sliding window length and a fixed sliding step size, pressure and flow response segments are extracted from the quality control aligned flow set, and multi-source aligned segments within the same window are extracted simultaneously to generate a multi-source aligned response window sequence.

3. The intelligent grouting calculation method based on deep neural networks as described in claim 2, characterized in that: The steps for generating the fracture inversion parameter set are as follows: The deep neural network for crack inversion includes an input embedding layer, a one-dimensional convolutional layer, a gated recurrent node layer, and a fully connected mapping layer; The multi-source aligned response window sequence is input into the input embedding layer of the crack inversion deep neural network, and channel rearrangement and window position encoding are performed to generate a unified channel sequence. Based on a unified channel sequence, texture features of abrupt segments and morphological features of steady-state segments are extracted by local convolution along the time axis in a one-dimensional convolutional layer. Based on the texture of the abrupt segment and the morphological features of the steady-state segment, cross-sampling point dependencies are aggregated in the gated loop node layer to generate a temporal representation vector; In the fully connected mapping layer, the temporal representation vector is inverted and projected to generate a fracture inversion parameter set.

4. The intelligent grouting calculation method based on deep neural networks as described in claim 3, characterized in that: The steps for partitioning and optimizing the fracture inversion parameter set are as follows: Based on the fracture inversion parameter set, the distal water-blocking zone and the proximal reinforcement zone are divided according to the axial segmentation rules of the grouting hole, and the hole segment parameters are collected and partitioned in combination with the hole segment identifier to generate the hole segment parameter queue. The parameters of the partitioned borehole segments are rearranged and the constraints are associated and organized according to the priority of water plugging and reinforcement, respectively, to generate a partitioned optimization parameter table.

5. The intelligent grouting calculation method based on deep neural networks as described in claim 4, characterized in that: The steps for generating the multi-level grouting arrangement input package are as follows: The zoning optimization parameter table and grouting process parameters are expanded by unifying the hole segment identification diameter and arranging the process fields to form a merged process parameter table. The process merging parameter table and construction constraints are matched with the hole segment identifier key and the constraint field is concatenated to generate a multi-level grouting arrangement input package.

6. The intelligent grouting calculation method based on deep neural networks as described in claim 5, characterized in that: The steps for generating the orchestration list in the generation phase are as follows: The grouting arrangement deep neural network includes a multi-head self-attention encoding layer, a stage boundary generation layer, a trajectory decoding layer, and a differentiable constraint projection layer; The multi-level grouting arrangement input package is input into the grouting arrangement deep neural network, and query key-value projection and attention convergence are performed in the multi-head self-attention encoding layer to generate a global arrangement representation. In the stage boundary generation layer, the global orchestration representation is deduced into differentiable stages and the stage boundaries are located to generate the stage sequence and stage duration. Based on the phase sequence and phase duration, the phase conditions are expanded and the sampling points are gradually decoded and derived through the trajectory decoding layer. Combined with the construction constraints, the grouting process parameters are optimized to generate pressure and flow trajectory sampling points. In the differentiable constraint projection layer, the pressure and flow trajectory sampling points are continuously projected and the boundaries are clipped to generate a stage orchestration list.

7. The intelligent grouting calculation method based on deep neural networks as described in claim 6, characterized in that: The process of performing staged grouting according to the staged arrangement list, and collecting measured pressure, measured flow rate, and backflow volume, is as follows: The pressure and flow trajectory sampling points are expanded into a control cycle execution sequence according to the phased arrangement list, and the grouting pump control interface is issued. At the same time, the control cycle number is written to form a phased execution instruction. During the control cycle, the measured pressure, measured flow rate, and backflow volume are collected synchronously and aligned according to the control cycle number to generate a three-quantity aligned sampling entry.

8. The intelligent grouting calculation method based on deep neural networks as described in claim 7, characterized in that: The steps for generating the intelligent grouting calculation record are as follows: The arrival time difference elimination and missing measurement anomaly handling are performed on the three-quantity aligned sampling entries, and the results are collected into a phase execution receipt sequence; Align the phase execution receipt sequence with the phase orchestration list according to the control cycle number, aggregate the actual execution trajectory of the phase, organize it into an execution receipt response window sequence, and encapsulate it into a grouting intelligent calculation record.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the grouting intelligent calculation method based on deep neural networks as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the grouting intelligent calculation method based on deep neural networks as described in any one of claims 1 to 8.

Images