Parameter correlation neural network-based filling proportion adaptive optimization method

Abstract

The present application belongs to the technical field of filling process parameter optimization, and relates to a filling proportioning adaptive optimization method based on a parameter correlation neural network, comprising: acquiring continuously collected rheological shear force signals and microseismic acoustic wave signals to generate synchronous physical signals; analyzing the high-frequency envelope characteristics of the signals to extract transient phase derivatives, combining an interference phase change threshold trigger node mapping rule to generate a connection index sequence; analyzing the sequence to reconstruct material node correlation, generating a dynamic topology adjacency matrix, injecting the matrix reconstruction operation path and instantiating to generate an instance graph neural network; calling the execution of inference calculation, outputting the intensity prediction residual value and the comprehensive cost evaluation vector, and executing target optimization in cooperation with both to output the best proportioning instruction. The present application solves the problem that the existing technology is difficult to capture the microscopic interference state of multi-component material mixing, resulting in that the evaluation model is not easy to update synchronously with the real physical coupling relationship.

Description

Technical Field

[0001] This invention belongs to the technical field of filling process parameter optimization, and relates to an adaptive optimization method for filling ratio based on parameter correlation neural network. Background Technology

[0002] Mine backfilling is the process of mixing tailings, cement, and other cementing materials with water in a specific ratio to form a backfill slurry, which is then transported underground. The backfill mix proportion determines the mechanical strength, rheological transport characteristics, and overall cost of the backfill. During production, the physicochemical properties of tailings fluctuate due to changes in the mining environment, and parameters such as particle size and mineral composition exhibit dynamic changes. The variability in the characteristics of incoming materials increases the complexity of maintaining backfill quality and controlling costs.

[0003] Currently, commonly used mix design optimization methods are mainly divided into two categories. The first category is the table-lookup decision method, where technicians develop a decision table that meets the design strength through sampling analysis and mix design experiments, and select a fixed formula on-site based on tailings classification. The second category is the traditional mathematical modeling method, which establishes a linear or polynomial regression model based on historical data to fit the mapping relationship between strength and mix design, and uses a planning algorithm to find the lower-cost solution while meeting the strength requirements.

[0004] The aforementioned lookup-based decision-making method has a relatively long experimental cycle, and the analysis results are prone to a time lag with the actual production progress. This makes it difficult for the proportion adjustment to respond promptly to dynamic changes in material properties, and in scenarios with fluctuating incoming materials, proportion mismatches are likely to occur, thus affecting the strength compliance rate of the filling body or increasing material consumption. Meanwhile, traditional mathematical modeling methods are mostly based on linear system assumptions, and their model structures are relatively simple, making it difficult to accurately characterize the nonlinear coupling effects caused by particle collisions and shear thinning during the mixing of multi-component materials. When the number of material components increases, the prediction accuracy of the model decreases, and its optimization results do not adequately express the physical interference state, limiting its ability to analyze and predict macroscopic intensity change trends. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides an adaptive optimization method for filling ratio based on a parameter-related neural network.

[0006] The adaptive optimization method for filling ratio based on parameter correlation neural network includes the following steps: S1. Acquire the rheological shear force signal and micro-vibration acoustic wave signal continuously collected by the detection device, and perform signal alignment based on time delay compensation to generate a synchronous physical signal; S2. Analyze the high-frequency envelope characteristics of the synchronous physical signal and extract the transient phase derivative; S3. Based on the comparison results of transient phase derivative and interference phase transition threshold, the node mapping rule is triggered to generate a node connection index sequence; S4. Parse the node connection index sequence to reconstruct the material node association and generate a dynamic topological adjacency matrix; S5. Inject the dynamic topological adjacency matrix to reconstruct the computation path of the graph nodes, instantiate the initial graph neural network, and generate the instance graph neural network; S6. Call the instance graph neural network to perform forward graph convolutional inference calculation, decode the output intensity prediction residual value and the comprehensive cost evaluation vector; S7. Combine the intensity prediction residual value and the comprehensive cost assessment vector to perform multi-level target penalty optimization and output the target filling ratio instruction.

[0007] A further aspect of the present invention, step S1, includes the following steps: Extract the inter-channel time deviation between the rheological shear force signal and the micro-vibration acoustic wave signal; The cross-correlation integral algorithm is used to process rheological shear force signals and micro-vibration acoustic signals to quantify the time delay between channels. The rheological shear force signal is optimized by time axis translation and phase compensation using the time delay between channels, and then aligned with the micro-vibration acoustic wave signal to generate a synchronous physical signal.

[0008] A further aspect of the present invention, step S2, includes the following steps: Transient intrinsic mode decomposition is performed on the data streams of each communication channel of the synchronous physical signal to remove the macroscopic steady-state low-frequency components and separate the high-frequency interference envelope with a frequency higher than the preset center frequency threshold. The analytical signals of the high-frequency interference envelope sequence corresponding to multi-component materials are calculated by Hilbert transform to generate dynamic cross-analytical signals and capture the extreme point information of the energy envelope of the signal. Based on the extreme point information, the first-order time difference value of the cross-phase function is obtained, and the transient phase derivative is calculated.

[0009] A further aspect of the present invention, step S3, includes the following steps: Determine whether the transient phase derivative crosses the preset interference phase transition threshold, and convert the transient phase derivative that has passed the threshold screening into a dynamic step size factor; Based on the dynamic step size factor, the parameter node state in the preset one-dimensional node state array is updated, and the target node whose state identifier is updated to the active state is obtained. Collect the array index features of the target nodes and convert them into node index offsets. Then, concatenate the continuously generated node index offsets to construct a node connection index sequence.

[0010] A further aspect of the present invention, step S4, includes the following steps: Parse and extract the consecutive index instructions contained in the node connection index sequence; The continuous indexing instructions are projected onto the preset two-dimensional node association space that represents the distribution of material types, and the edge connection weights between material nodes are initially determined. Remove invalid coupled lines in the two-dimensional node association space that do not match the coordinate position corresponding to the index command, and lock and retain valid association edges; Based on effective associated edges and preset active material nodes, a node connectivity graph is constructed, and a dynamic topological adjacency matrix is ​​encapsulated to adapt to transient mixed features.

[0011] A further aspect of the present invention, step S5, includes the following steps: Initialize and create an initial graph neural network containing multiple graph nodes; The dynamic topological adjacency matrix is ​​extracted as a shape constraint that defines the dimension of the convolution tensor of the underlying graph; Extract shape constraints to update the empty connection architecture of the initial graph neural network; Based on shape constraints, feature computation channels are assigned to graph node pairs that establish connectivity, and feature propagation masking is performed on graph node pairs that do not have connectivity, thus generating instance graph neural networks.

[0012] A further aspect of the present invention, step S6, includes the following steps: Obtain the standard material ratio baseline value and candidate adjustment range variables required for the target processing cycle; The standard material ratio baseline value and candidate adjustment range variables are fused and transmitted to the input layer of the instance graph neural network; The forward graph convolution extraction and activation function propagation are performed sequentially along the node connection channels of the instance graph neural network, and a preset validation set distribution threshold is applied to suppress feature value overflow. The multidimensional convergence tensor generated by the output layer of the aggregated instance graph neural network is used for feature dimensionality reduction, separation, decoding, and output of intensity prediction residuals and comprehensive cost evaluation vectors.

[0013] A further aspect of the present invention, step S7, includes the following steps: Different convergence weight coefficients are preset for the intensity prediction residual value and the comprehensive cost assessment vector, and a fitness assessment function containing a cost penalty term is synthesized. Configure joint constraint boundary conditions to maximize the static pass rate of the filling body and minimize the cost of filling material per unit volume; The nonlinear heuristic algorithm is activated to calculate the fitness evaluation function in real time within the preset ratio adjustment range space, and to search for the target solution set containing the target adjustment range variable; Numerical target parameters are extracted from the target solution set, jointly calculated, and converted into target filling ratio instructions conforming to the communication protocol, which are then sent to the field control layer.

[0014] A further aspect of the present invention includes, before performing cross-correlation integration on the rheological shear force signal and the micro-vibration acoustic signal, the following: The rheological shear force signal and the micro-vibration acoustic signal are subjected to noise reduction preprocessing using a preset mean filtering algorithm to eliminate background noise interference. The absolute physical units carried by the data stream after noise reduction and preprocessing are removed by the zero-mean normalization equation, and a dimensionless scalar function matrix is ​​generated. The dimensionless scalar function matrix is ​​used as the recombined rheological shear force signal and the microseismic acoustic signal to perform cross-correlation integration, thus shielding cross-dimensional calculation errors.

[0015] A further aspect of this invention, after extracting numerical target parameters from the target solution set, jointly calculating them, converting them into target filling ratio instructions conforming to the communication protocol, and sending them to the field control layer, further includes: Simultaneously initiate the data retrieval task to receive the actual strength measurement parameters obtained from the standard filling test blocks formed by on-site physical execution after standard curing; Calculate the numerical deviation term between the actual intensity measurement parameters and the intensity value predicted by the network model in the current execution cycle; The actual intensity measurement parameters are used as supervision labels and substituted into the loss function to calculate the gradient, driving the online incremental training of the weights of the bottom convolutional layers of the instance graph neural network. At the same time, the numerical deviation term is extracted as an error compensation correction factor to adjust the interference phase transition threshold.

[0016] In summary, the present invention has the following beneficial technical effects: 1. This invention quantifies the inter-channel time delay between rheological shear force signals and micro-vibration acoustic signals by performing cross-correlation integration, and then uses this delay to perform time axis translation and phase compensation optimization on the original signals. This technical solution can align multi-channel signals that are misaligned due to differences in the physical response mechanisms of the detection devices under a unified time reference, reducing waveform misalignment errors caused by differences in sensor hardware circuit delays. This results in better time consistency of the physical features extracted from the multi-source signals, which helps improve the timing accuracy of the joint identification of micro-rheological interference events using multi-source signals. 2. This invention controls the dynamic step size factor by comparing the transient phase derivative with a preset interference phase transition threshold, thereby driving the state update of the target node in the node state array. This mechanism establishes a mapping relationship between the transient change characteristics of high-frequency physical signals and the activation instructions of underlying computing resources, enabling the system to allocate subsequent computing resources for parameter combinations that exhibit physical interlocking effects at the current moment. This reduces the occupation of computing channels by parameter combinations without actual physical coupling, and improves the computational targeting of the neural network during topology reconstruction. 3. This invention constructs a dynamic topological adjacency matrix adapted to the current transient mixing characteristics by projecting the node connection index sequence onto a two-dimensional node association space, retaining valid association edges matching the activation command and eliminating invalid connections. This scheme utilizes actual physical interference characteristics to define the information transmission paths between nodes in the graph neural network, reducing the invalid coupling phenomena that may be introduced by a preset fixed network topology model. This allows the model structure to dynamically evolve with changes in the rheological state of the material, improving the model's ability to represent complex physical variations during the mixing of multi-component materials.

[0017] 4. This invention synthesizes a fitness evaluation function with a cost penalty term by configuring different convergence weight coefficients for the intensity prediction residual value and the comprehensive cost evaluation vector, and then uses a nonlinear heuristic algorithm for optimization under joint constraint boundary conditions. This optimization strategy can dynamically adjust the decision-making tendency according to real-time safety standards and economic expectations. When potential intensity risks are assessed, corresponding penalty weights are added, and economic costs are taken into account when the system is running smoothly. This achieves a dynamic balance between filling safety and economic benefits in multi-objective decision-making and reduces the reliance on manually set fixed priority rules. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings are used to provide a further understanding of the present invention.

[0019] Figure 1 A flowchart illustrating an embodiment of this application is disclosed.

[0020] Figure 2 The cross-correlation coefficients versus time offset curves in the embodiments of this application are disclosed.

[0021] Figure 3 Structural schematic diagrams of embodiments of this application are disclosed. Detailed Implementation

[0022] The following is in conjunction with the appendix Figures 1-3 A preferred description of the present invention is provided below.

[0023] See attached document Figure 1 This invention proposes an adaptive optimization method for filling ratio based on a parameter correlation neural network, comprising the following steps: S1. Acquire the rheological shear force signal and micro-vibration acoustic wave signal continuously collected by the detection device, and perform signal alignment based on time delay compensation to generate a synchronous physical signal; S2. Analyze the high-frequency envelope characteristics of the synchronous physical signal and extract the transient phase derivative; S3. Based on the comparison results of transient phase derivative and interference phase transition threshold, the node mapping rule is triggered to generate a node connection index sequence; S4. Parse the node connection index sequence to reconstruct the material node association and generate a dynamic topological adjacency matrix; S5. Inject the dynamic topological adjacency matrix to reconstruct the computation path of the graph nodes, instantiate the initial graph neural network, and generate the instance graph neural network; S6. Call the instance graph neural network to perform forward graph convolutional inference calculation, decode the output intensity prediction residual value and the comprehensive cost evaluation vector; S7. Combine the intensity prediction residual value and the comprehensive cost assessment vector to perform multi-level target penalty optimization and output the target filling ratio instruction.

[0024] In one embodiment of the present invention, step S1 includes the following steps: Extract the inter-channel time deviation between the rheological shear force signal and the micro-vibration acoustic wave signal; The cross-correlation integral algorithm is used to process rheological shear force signals and micro-vibration acoustic signals to quantify the time delay between channels. The rheological shear force signal is optimized by time axis translation and phase compensation using the time delay between channels, and then aligned with the micro-vibration acoustic wave signal to generate a synchronous physical signal.

[0025] Specifically, the main control processing node activates the detection devices deployed on the filling and mixing chamber nodes through the system's underlying bus. The detection devices continuously contact the tailings slurry in the mixing field and continuously collect rheological shear force signals and micro-vibration acoustic wave signals that reflect the rheological process of collision and mixing of multi-component materials.

[0026] Due to inherent differences in the hardware conversion response mechanisms of the two types of sensors, asynchronous stray delays exist between synchronously accessed signal channels. The system extracts the inter-channel time deviation caused by the physical response delay of the detection devices, and uses a cross-correlation integral algorithm to process the rheological shear force signal and the micro-vibration acoustic wave signal after noise reduction preprocessing. By calculating the similarity peak position of the two independent time-domain sequences through a sliding time difference window, the inter-channel time delay of the rheological shear force signal and the micro-vibration acoustic wave signal can be accurately quantified.

[0027] After obtaining the inter-channel time delay, the system uses the clock stamp of the micro-vibration acoustic signal as the absolute stationary reference. It then performs time-axis translation on the data sequence of the rheological shear force signal using the inter-channel time delay, and reconstructs missing data points using an interpolation algorithm. This aligns and merges the misaligned waveform signals on the time axis, outputting a synchronous physical signal to provide a time-domain input source for downstream analytical steps. Cross-correlation integration is then performed to calculate the inter-channel time delay.

[0028]

[0029] In the formula, This represents the rheological shear force signal after feature scaling. This represents a micro-vibration acoustic signal that has undergone the same processing. This represents the length of the integration window preset by the system. This represents the time shift offset variable during the sliding window process of integration. This represents the output sequence of consecutive cross-correlation coefficients.

[0030] In practical engineering calculations, for For potential negative time indices, the system extends and covers them by calling prior historical data stored in the circular buffer to ensure the continuity and mathematical rigor of the integration operation in the time domain.

[0031] Find the cross-correlation function using an optimization algorithm. Time offset when the global maximum value is reached Then This is assigned and set as the inter-channel time delay. In digital signal processing systems, the above continuous cross-correlation integral operation is equivalently implemented by performing a sliding summation on the discrete time series.

[0032] The filling and mixing chamber node refers to the three-dimensional mechanical cavity area in a modern underground mine filling preparation plant responsible for the forceful mixing of cementitious materials and multiphase gangue sand. The detection device consists of a strain gauge array mounted at the root of the mixing blades and a broadband piezoelectric acoustic emission array attached to the outer wall of the cavity.

[0033] Regarding the length of the integration window The settings assume an application scenario of high-concentration, high-flow-rate tailings backfilling system. In this case, the mixture exhibits strong non-Newtonian shear thickening characteristics, resulting in a relatively long physical period for rheological interference scouring. The integral window length is set based on the empirical average rotational speed of the mixing spindle. The duration is three to five times the numerical duration of a single complete rotation cycle of the main axis to eliminate spurious correlation peaks caused by abrupt noise.

[0034] In one example of this invention, the system sends a high-level on-detection device to trigger data stream injection. Based on the on-site stirring spindle speed of 20 rpm, the system extracts the length of the integration window. The sample array is 9s long. At this length, the buffer register retains and truncates a high-dimensional rheological shear force signal array and a micro-vibration acoustic wave signal array with a dimension of 900. Before execution, the unit performs subtraction to eliminate the mean of each sample and divides by the standard deviation to perform dimensionless scaling. Then, with a time shift offset step of 2ms, cumulative multiplication is successively superimposed to calculate the cross-correlation coefficient sequence. See also Figure 2 The calculation and deduction show that when the shift variable τ is 48ms, the cross-correlation coefficient is... The extreme value of 0.85 was obtained. Based on this, the system determined the time delay between channels caused by the inherent deviation in physical transmission between the two to be 48ms. According to this time, the control algorithm shifted the time series address of the rheological shear force signal stored in the buffer stack forward by 48ms, and used the cubic spline interpolation function to reset the phase angle gap caused by the shift to perform phase compensation optimization, reducing the alignment error to within the 0.5ms tolerance threshold. The two data were then sent in parallel to generate a synchronous physical signal to wait for the subsequent high-frequency reconstruction and decoding stage.

[0035] In one embodiment of the present invention, step S2 includes the following steps: Transient intrinsic mode decomposition is performed on the data streams of each communication channel of the synchronous physical signal to remove the macroscopic steady-state low-frequency components and separate the high-frequency interference envelope with a frequency higher than the preset center frequency threshold. The analytical signals of the high-frequency interference envelope sequence corresponding to multi-component materials are calculated by Hilbert transform to generate dynamic cross-analytical signals and capture the extreme point information of the energy envelope of the signal. Based on the extreme point information, the first-order time difference value of the cross-phase function is obtained, and the transient phase derivative is calculated.

[0036] The system invokes the signal processing unit to perform transient intrinsic mode decomposition (EMD) on the data streams of the rheological shear force channel and the micro-vibration acoustic wave channel of the synchronous physical signal in parallel. This operation uses an empirical mode decomposition algorithm or a variant thereof to adaptively decompose the composite signal of each channel into a set of intrinsic mode function components whose amplitude and frequency vary with time, thereby stripping away the macroscopic steady-state low-frequency components and separating the high-frequency interference envelope with a frequency higher than a preset center frequency threshold.

[0037] The calculation module locks the high-frequency interference envelope sequence of the corresponding channel for different materials, obtains its analytical signal through Hilbert transform, and calculates the dynamic cross-analytic signal of the two analytical signals to extract the cross-envelope extreme point information corresponding to the high coupling of signal energy in the time domain when physical interference occurs.

[0038] Based on the precise timestamp corresponding to the extreme point information, the system extracts the cross-phase function within a preset time window neighborhood of this timestamp and performs phase unwinding processing on it to eliminate... After the numerical singularity generated by the phase jump, its first-order time difference is solved, and the scalar value is calculated using numerical differentiation methods; this scalar value is the transient phase derivative. The formula for solving the transient phase derivative is:

[0039]

[0040] High-frequency interference envelope sequence obtained by Hilbert transform and The analytical signals are multiplied to obtain the dynamic cross-analytic signal. The system extracts its actual part. With the imaginary part Calculate the transient cross phase function It performs phase unwinding to ensure phase continuity on the time axis.

[0041] Subsequently, at the time point determined by the cross-envelope extreme point information At this point, the transient cross-phase function is calculated using the central difference method. The first-order time derivative at this point yields the transient phase derivative. .in This is the preset time step.

[0042] Transient intrinsic mode decomposition (TEM) specifically refers to an adaptive decomposition method that can handle non-stationary and nonlinear signals. The system defaults to using the ensemble empirical mode decomposition algorithm, and its selection and iteration stopping criteria are set based on the standard deviation threshold. The typical value of this threshold is usually between 0.2 and 0.3 to balance the thoroughness of the decomposition and the computational efficiency.

[0043] The high-frequency interference envelope refers to the component among the several intrinsic mode functions obtained after decomposition whose center frequency is higher than a preset macroscopic steady-state characteristic frequency threshold of the fluid. This threshold is set based on statistical analysis of historical operating data and is usually 50-100 times the rotation frequency of the stirring spindle. Cross-envelope extreme point information refers to the timestamp information corresponding to local peak points on the transient amplitude envelope of the dynamic cross-analytical signal whose amplitude exceeds three standard deviations of the global mean.

[0044] In one example of the present invention, the system receives the synchronous physical signal output from step S1, the data format of which is two aligned time series of length 900. The system performs ensemble empirical mode decomposition on these two sequences, separating the first three eigenmode functions of each as high-frequency interference envelopes. Subsequently, the calculation module performs Hilbert transform on the first mode function of the rheological shear force channel and the first mode function of the micro-vibration acoustic wave channel and obtains the dynamic cross-analytic signal, identifying the timestamp in their amplitude envelopes. A significant energy peak was observed at this point, which was identified as the extreme point of the cross-envelope. The system further extracted the value of the transient cross-phase function near this time point and queried the value at that time. The time phase value is ,exist The time phase value is Set time step The value is 0.01s. Substituting this into the difference formula, the transient phase derivative is calculated as follows: This value will be passed to downstream modules as a trigger constraint key.

[0045] In one embodiment of the present invention, step S3 includes the following steps: Determine whether the transient phase derivative crosses the preset interference phase transition threshold, and convert the transient phase derivative that has passed the threshold screening into a dynamic step size factor; Based on the dynamic step size factor, the parameter node state in the preset one-dimensional node state array is updated, and the target node whose state identifier is updated to the active state is obtained. Collect the array index features of the target nodes and convert them into node index offsets. Then, concatenate the continuously generated node index offsets to construct a node connection index sequence.

[0046] The system will receive the transient phase derivative from the previous processing step as an input signal and use the value of this signal to trigger the underlying memory addressing rules.

[0047] First, the system compares the absolute value of the transient phase derivative with a preset interference phase transition threshold to determine whether the current degree of microscopic interference of the material has reached a critical state sufficient to cause a significant change in macroscopic properties. If the absolute value of the transient phase derivative is greater than the interference phase transition threshold, the system converts the specific value of the derivative into a dynamic step size factor through a preset nonlinear mapping function or lookup table.

[0048] The pre-defined nonlinear mapping function or lookup table is based on the physical characteristic that different material components, due to differences in particle size distribution and specific surface area, will generate phase derivatives with specific frequency bands and slope ranges during rheological interlocking and flocculation. This achieves a unique mapping from phase derivative values ​​to specific physical interlocking events. Subsequently, the system performs write operations on a pre-defined one-dimensional node state array, which characterizes the coupling relationships between parameters, based on this dynamic step size factor. Specifically, the dynamic step size factor is used as an index or pointer to precisely locate one or more target nodes in the one-dimensional node state array and drive the state identifiers of these located target nodes to update from an initial inactive state to an active state, thereby marking the parameter associations activated by the physical interference process at the current moment.

[0049] After a complete computation cycle, the system controller scans the entire one-dimensional node state array, collecting the array index features (i.e., array index values) of all active target nodes, and directly converts these index values ​​into node index offsets. Finally, the system concatenates multiple consecutively generated node index offsets according to their generation time order to construct and generate an ordered integer sequence, which is the node connection index sequence. This sequence will serve as the direct instruction for reconstructing the network topology in the next stage.

[0050] The interference phase transition threshold is a crucial triggering threshold. Its physical significance lies in defining the critical point at which the microstructure within a material evolves from linear, reversible perturbation to nonlinear, irreversible flocculation or agglomeration. This threshold is determined based on statistical analysis of extensive historical rheological experimental data of filling slurries. It is identified by determining the minimum rate of phase change that causes an irreversible increase in slurry viscosity at different shear rates. When assuming the application scenario involves processing complex tailings with high mud content, this value is typically preferred to be in the range of 10.0 rad / s to 15.0 rad / s.

[0051] The one-dimensional node state array is a predefined one-dimensional Boolean array in memory. Its length is equal to the total number of pairwise combinations between all material components and process parameters under investigation. Each index position in the array is hard-coded to uniquely map to a specific parameter interaction pair; for example, index 0 represents the interaction between cement and water, and index 1 represents the interaction between water and fine tailings. The inactive and active state flags are typically set to integers 0 and 1 respectively at the digital logic level to represent the disconnected and connected states of the parameter association channels.

[0052] In one example of this invention, the transient phase derivative value acquired by the system is 16.0 rad / s. The system retrieves a preset interference phase transition threshold from the configuration library, which is 12.0 rad / s. The system performs a comparison operation. Since the absolute value of the transient phase derivative, 16.0 rad / s, is greater than the interference phase transition threshold of 12.0 rad / s, the system determines that a significant physical interference event has occurred and triggers subsequent addressing logic. The system further inputs the value of 16.0 rad / s into a preset mapping rule module. This module, according to internally defined mapping rules reflecting the collision characteristics of material particles, for example, maps phase derivative values ​​in the range of 15.0 rad / s to 20.0 rad / s, representing a violent flocculation reaction between two types of materials, to index 5, thereby determining that the event corresponds to the target node with index 5 in the one-dimensional node state array. Assume that the initial state of the array is a vector of all zeros. Based on this, the system updates the element at index 5 from the inactive state flag 0 to the active state flag 1, thus updating the state of the one-dimensional node state array to... At the end of the current computation cycle, the system scans the array and collects the set of active node indices as {5}. The index value 5 is directly adopted as the node index offset. Therefore, the system finally splices and constructs a sequence of node connection indices containing only a single element [5], and outputs it to the downstream topology reconstruction module.

[0053] In one embodiment of the present invention, step S4 includes the following steps: Parse and extract the consecutive index instructions contained in the node connection index sequence; The continuous indexing instructions are projected onto the preset two-dimensional node association space that represents the distribution of material types, and the edge connection weights between material nodes are initially determined. Remove invalid coupled lines in the two-dimensional node association space that do not match the coordinate position corresponding to the index command, and lock and retain valid association edges; Based on effective associated edges and preset active material nodes, a node connectivity graph is constructed, and a dynamic topological adjacency matrix is ​​encapsulated to adapt to transient mixed features.

[0054] Specifically, the system first parses the node connection index sequence received from the previous calculation step. This sequence contains one or more integer indices representing the activated parameter association channels. Simultaneously, the system initializes a two-dimensional node association space in memory representing the complete distribution of material types. Its physical form is a space with dimensions of [missing information]. The square array, in which To fill the preset total number of nodes in the graph, all elements of the square matrix are initially set to zero. The system then converts each integer index in the node connection index sequence into a unique coordinate pair in the two-dimensional node association space using a predefined mapping function from one-dimensional indices to two-dimensional matrix coordinates. .

[0055] This coordinate pair and its symmetric coordinates The corresponding matrix element values ​​are modified from zero to one to initially determine the edge connection weights between material nodes. During this process, the element values ​​of all coordinate positions in the two-dimensional node association space that are not indexed by the node connection index sequence remain zero, which is equivalent to logically eliminating all invalid coupling connections that do not match the index instructions. Through this selective bit-setting operation, the system retains only valid association edges representing effective physical interference characteristics. Finally, the system integrates all valid association edges with the node connectivity graph of all corresponding active material nodes, directly encapsulates the sparse matrix formed in the current state, constructs and outputs a dynamic topological adjacency matrix specifically adapted to the current transient mixed characteristics, serving as the sole basis for downstream neural network topology reconstruction.

[0056] Define a dynamic topological adjacency matrix elements Generation rules:

[0057] In the formula, It is a set consisting of all the indices in the node-connected index sequence. It is a predefined inverse mapping function used to associate coordinate pairs of two-dimensional nodes in space. Transform back to its unique index in the one-dimensional node state array if and only if the coordinate pair or its symmetric coordinates The corresponding one-dimensional index exists in the set In the middle, matrix elements The value is assigned to 1, indicating that the node With nodes There is a bidirectionally connected edge between them; in all other cases, the element is assigned a value of 0.

[0058] It should be noted that the two-dimensional node association space is a logically fully connected graph, which is usually stored in the form of an adjacency matrix in computers, and its dimension is... The number of material components and key process parameters that need to be monitored and optimized in the process is determined by the total number of such parameters. Active material nodes correspond to the parameters represented by each row and column of the adjacency matrix, such as cement dosage, tailings particle size, and stirring speed. Effective connections are represented by elements with a value of 1 in the matrix, indicating a strong coupling relationship between the corresponding two parameter nodes under the current physical rheological state, determined by the underlying physical mechanism. Conversely, invalid coupling connections are represented by elements with a value of 0, indicating that there is no direct and significant mutual influence between the corresponding two parameter nodes at this moment.

[0059] In one example of this invention, firstly, according to the process design, a 5×5 two-dimensional node association space is initialized, representing five active material nodes. This space is initially an all-zero matrix. The system has a pre-defined mapping rule from one-dimensional indices to two-dimensional coordinate pairs, which uniquely maps index 5 to the node pair (1, 3). The system parses the node connection index sequence and extracts the index value 5. Based on this index value and the mapping rule, the system sets the values ​​of the elements with coordinates (1, 3) and their symmetric coordinates (3, 1) in the two-dimensional node association space from 0 to 1. After this operation, only two elements in the 5×5 matrix are 1, while the remaining 23 elements are still 0. This is equivalent to preserving the valid association edge between node 1 and node 3, while eliminating all other possible invalid coupling connections. Finally, the system encapsulates this highly sparse 5×5 matrix and outputs a dynamic topological adjacency matrix. The specific form of this matrix is:

[0060] This matrix will serve as a mandatory constraint for the generation of the graph neural network structure in the next stage.

[0061] In one embodiment of the present invention, step S5 includes the following steps: Initialize and create an initial graph neural network containing multiple graph nodes; The dynamic topological adjacency matrix is ​​extracted as a shape constraint that defines the dimension of the convolution tensor of the underlying graph; Extract shape constraints to update the empty connection architecture of the initial graph neural network; Based on shape constraints, feature computation channels are assigned to graph node pairs that establish connectivity, and feature propagation masking is performed on graph node pairs that do not have connectivity, thus generating instance graph neural networks.

[0062] Specifically, the system first performs an initialization setup operation in memory, creating one or more initial graph neural networks. In the initial state, although the network defines graph nodes representing different filling parameters, the adjacency matrix used to describe the connections between nodes is set to empty or all zeros. Therefore, the communication channels between nodes in its hidden layers are in a logically disconnected state where feature calculation channels are not allocated, and effective graph convolution operations cannot be performed.

[0063] The system extracts and updates the dynamic topological adjacency matrix generated in the previous processing step into the network model, specifying it as the shape constraint defining the dimensions of the underlying graph convolutional tensors. This update operation updates the empty connection architecture of the initial graph neural network, using the sparse structure of the dynamic topological adjacency matrix as the sole topological basis for message passing between network layers. Finally, the system initiates the instantiation process according to this shape constraint.

[0064] For graph node pairs with connectivity in the dynamic topological adjacency matrix, the system assigns them feature computation channels with actual multiply-accumulate capabilities, constructing effective data interaction paths. Simultaneously, for all unconnected graph node pairs, the system performs feature pass-through masking, directly ignoring these potential connections during computation graph compilation, thus forcibly zeroing out their corresponding gradient data flows. Through this series of operations, the system instantiates and generates instance graph neural networks that match the current actual physical mechanism, for direct use in forward inference.

[0065]

[0066] This formula represents the core operation of the forward propagation in a graph convolutional layer, demonstrating how a dynamic topological adjacency matrix reconstructs the computation path. Specifically, It is the first The feature matrix of all graph nodes in the layer, This is the trainable weight matrix for this layer. Core variables. The adjacency matrix with self-loops injected, i.e. , It is the identity matrix. This is the dynamic topological adjacency matrix output from the previous implementation step and injected here as a shape constraint. It is based on the matrix The calculated degree matrix. This represents a non-linear activation function. This formula shows that introducing the identity matrix preserves the historical characteristics of the nodes themselves, and only when the dynamic topological adjacency matrix... Corresponding element When non-zero, node and nodes Only when the features of these features are aggregated during the convolution process can dynamic network structures driven by physical phenomena be constructed.

[0067] An initial graph neural network refers to a neural network object where the number of nodes and the initial feature dimensions of the nodes are predefined at the software level, but its adjacency matrix is ​​not filled, resulting in an incomplete graph structure. Feature computation paths are dynamically allocated memory buffers and computational core scheduling paths used by graphics processing units (GPUs) or central processing units (CPUs) to perform sparse-to-dense matrix multiplication.

[0068] Dead zone blocking is a computation optimization strategy that, when constructing a computation graph, does not generate corresponding multiplication or addition instructions for zero elements in the adjacency matrix, thereby avoiding invalid computations and related memory accesses. The direct result is that during backpropagation, the gradient values ​​on these blocked paths are zero.

[0069] In one example of this invention, the system initializes an initial graph neural network containing five nodes, representing cement, water, coarse tailings, fine tailings, and additives, respectively. At this point, these five nodes are logically completely isolated from each other. Subsequently, the system uses the dynamic topological adjacency matrix generated in the previous embodiment... As the sole shape constraint, this constraint is forcibly injected into the network. During instantiation, the graph computation framework parses the matrix and finds non-zero values ​​only at coordinates (1, 3) and (3, 1). Therefore, the system allocates a feature computation path only between node 1, representing "water," and node 3, representing "fine tailings." For all other potential connections, such as between cement and water, feature propagation masking is performed. Ultimately, the system generates an instance graph neural network whose internal graph convolution operation path is uniquely determined: bidirectional feature information aggregation is performed only between the water node and the fine tailings node. The network is now ready to receive input data for inference.

[0070] In one embodiment of the present invention, step S6 includes the following steps: Obtain the standard material ratio baseline value and candidate adjustment range variables required for the target processing cycle; The standard material ratio baseline value and candidate adjustment range variables are fused and transmitted to the input layer of the instance graph neural network; The forward graph convolution extraction and activation function propagation are performed sequentially along the node connection channels of the instance graph neural network, and a preset validation set distribution threshold is applied to suppress feature value overflow. The multidimensional convergence tensor generated by the output layer of the aggregated instance graph neural network is used for feature dimensionality reduction, separation, decoding, and output of intensity prediction residuals and comprehensive cost evaluation vectors.

[0071] Specifically, the main control unit obtains the standard material ratio baseline value and candidate adjustment range variables required for the target processing cycle from the system task queue. The system adds these two sets of vectors element-wise to generate an initial node feature matrix, and then transmits it sequentially to the input layer of the instance graph neural network.

[0072] The network begins to perform forward graph convolutional inference computation. The data flow strictly follows the node connectivity channels that were effectively activated and mapped by the dynamic topological adjacency matrix in the previous instantiation step. Feature extraction and nonlinear activation function propagation are performed sequentially between the various graph convolutional layers of the network.

[0073] After each activation function propagation, the system performs a constraint operation, comparing each element value of the node feature vector with a preset validation set distribution threshold. Boundary clamping is applied to any values ​​exceeding this threshold to suppress feature value overflow. After propagation through all graph convolutional layers, the system aggregates the multi-dimensional node features of the output layer at the end of the instance graph neural network, using a global average pooling algorithm to reduce the dimensionality of the multi-node multi-dimensional aggregated tensor into a single graph-level feature vector.

[0074] Finally, the graph-level feature vector is fed into one or more fully connected decoding layers. After linear transformation and dimensionality reduction decoding, a scalar value is output, namely the strength prediction residual value that characterizes the final mechanical properties of the filling molding, and the comprehensive cost evaluation vector directly derived from the current material combination rules.

[0075]

[0076]

[0077]

[0078] The above formula describes the decoding process of the final output: First, by decoding the last graph convolutional layer... Output node feature matrix all node feature vectors Perform global average pooling to compute the graph-level aggregated feature vector. The aggregated feature vector is passed through a fully connected layer (with weights of ). , bias is Decoding yields the intensity prediction residual value. At the same time, the system uses standard material proportioning benchmarks. and candidate adjustment magnitude variables Calculate the material usage vector for the current proportion, divide it by a constant factor of 1000 to perform dimensional conversion, converting its mass unit to tons (tons per cubic meter), and then compare it with the preset unit cost vector. The Hadamard product is performed to calculate the comprehensive cost assessment vector representing the economic cost of a single filling volume. .

[0079] The standard material proportion baseline value is a vector stored in the system database, representing the validated basic formulation. Candidate adjustment amplitude variables are a set of perturbations generated by the previous-level optimization algorithm to explore better proportion schemes. The validation set distribution threshold is a two-dimensional vector, defining the minimum and maximum values ​​allowed for node features during inference. Its values ​​are determined by statistically analyzing a large amount of historical validation set data and taking the 99% confidence interval boundaries for each feature dimension. The strength prediction residual value is defined as the difference between the model-predicted 28-day compressive strength of the filling and the design target strength. The comprehensive cost assessment vector is a vector corresponding one-to-one with material types, where each element represents the unit volume cost of the corresponding material at that proportion.

[0080] In one example of this invention, the system retrieves the standard material proportioning baseline value from the database, which is in vector form as follows: These correspond to the amounts of cement, water, coarse tailings, fine tailings, and additives, respectively, in units of... At the same time, it receives candidate adjustment magnitude variables from the previous optimizer. The sum of the two yields the initial node feature input for that period. The feature matrix is ​​fed into the instance graph neural network generated in the previous embodiment. During forward propagation, the data flow strictly follows a topology where connectivity exists only between nodes 1 and 3. After multiple layers of graph convolution and activation function computation, the node feature matrix of the final output layer is assumed to be... The system then proceeded to... Perform global average pooling to obtain graph-level aggregated feature vectors. This vector is fed into the decoding layer, assuming decoding weights. and bias After training, the intensity prediction residual value is calculated. This indicates that the predicted intensity is lower than the target value. At the same time, the system retrieves the preset unit cost vector. Yuan / ton, the current proportion of usage After converting the mass unit to tons, perform a Hadamard product operation with the cost vector to obtain the comprehensive cost assessment vector. Yuan / .

[0081] In one embodiment of the present invention, step S7 includes the following steps: Different convergence weight coefficients are preset for the intensity prediction residual value and the comprehensive cost assessment vector, and a fitness assessment function containing a cost penalty term is synthesized. Configure joint constraint boundary conditions to maximize the static pass rate of the filling body and minimize the cost of filling material per unit volume; The nonlinear heuristic algorithm is activated to calculate the fitness evaluation function in real time within the preset ratio adjustment range space, and to search for the target solution set containing the target adjustment range variable; Numerical target parameters are extracted from the target solution set, jointly calculated, and converted into target filling ratio instructions conforming to the communication protocol, which are then sent to the field control layer.

[0082] First, the strength prediction residual and the comprehensive cost assessment vector are assigned different convergence weight coefficients based on pre-defined on-site safety standards and economic expectations. The system then sums the weighted strength residual penalty term and the weighted total cost term to synthesize a fitness assessment function with a dynamically adjustable cost penalty term.

[0083] The system sets joint constraint boundary conditions for the global search. These boundary conditions explicitly require that, during the optimization process, the predicted value of the static pass rate of the filling body must be maximized, while the unit cost of filling material obtained by summing the comprehensive cost evaluation vectors must be minimized. Based on this evaluation function and constraint conditions, the system initiates a nonlinear heuristic algorithm, such as particle swarm optimization or simulated annealing, to iteratively search within a preset ratio adjustment range space. It evaluates the fitness evaluation function value corresponding to each candidate ratio solution in real time to search for and locate the target solution set containing one or more solutions.

[0084] The system extracts candidate adjustment range variables whose values ​​meet the target from the target solution set, merges them with the standard material proportioning benchmark value, and converts them into target filling proportioning instructions that match the communication protocol of the industrial field programmable logic controller. These instructions are then sent to the field control layer via industrial Ethernet to directly drive the adjustment of the current batching valve opening on each material feed pipeline.

[0085] The system synchronously initiates a data retrieval task, receiving the actual intensity measurement parameters obtained in the laboratory, storing these measured values ​​in the historical dataset, and driving the online incremental training of the weights of the bottom convolutional layers of the instance graph neural network; simultaneously, it extracts the numerical deviation term as an error compensation correction factor, and calls back to correct the interference phase transition threshold in step S3. The specific construction form of the fitness evaluation function is as follows:

[0086] In the formula, It is the residual value of the intensity prediction. It is a comprehensive cost assessment vector. and These are the different convergence weight coefficients assigned to the strength and cost terms, ensuring that the synthesized fitness value is a scalar with a unified dimension. The design objective of this function is to minimize the fitness value. When the predicted strength value is lower than the target value, i.e. When the value is negative, a secondary penalty is applied through the squared term; when the intensity meets or exceeds the target, the intensity penalty term is zero. The cost term is simply the sum of all elements of the comprehensive cost evaluation vector.

[0087] Different convergence weight coefficients are hyperparameters set according to actual engineering needs. In the initial stage where absolute safety in production is required, the intensity weight may be adjusted. Set far higher than the cost weight. This is to ensure that all recommended proportions prioritize meeting strength requirements.

[0088] Once the process is stable, the temperature can be adjusted appropriately. The value is sought to find a more economical solution. Nonlinear heuristic algorithms specifically refer to a class of global optimization algorithms that do not rely on gradient information, suitable for handling complex, non-convex optimization problems with multiple local objective solutions. The target filling ratio instruction is a data message conforming to specific industrial communication standards, such as Modbus TCP / IP or PROFINET, which contains the target flow rate or quality setpoint for each material.

[0089] In one example of the present invention, the system obtains the intensity prediction residual value. and comprehensive cost assessment vector The system reads the weighting coefficients from the configuration and sets those that emphasize security. and taking into account economic efficiency Substitute the values ​​into the formula to calculate the fitness value of the current ratio. The particle swarm optimization algorithm records this fitness value and generates a new set of candidate adjustment magnitude variables, such as... Then, the calculation process for S6 and S7 is repeated. After hundreds of iterations, the algorithm finds that when At that point, the fitness function reached its minimum value of 195.5. The system then extracted the target adjustment amplitude variable and generated the final allocation vector. The data was encoded into a PROFIBUS-DP protocol data frame and sent to the on-site batching control PLC to adjust the valve openings to achieve the target proportion. 28 days later, the laboratory test results for this batch of filling material showed the target strength. The model prediction value is the target intensity. The system records this The numerical deviation term is used as a supervision label to trigger the backpropagation of the weights in the graph neural network model. Simultaneously, this numerical deviation term is used as an error compensation correction factor to fine-tune the value of the interference phase transition threshold. Specifically, when the measured intensity is higher than the model's predicted intensity, it indicates that the currently set interference phase transition threshold is too sensitive and introduces microscopic noise coupling edges. The system increases the threshold in the positive direction, for example, by fine-tuning it from 12.0 rad / s to 12.05 rad / s; conversely, it decreases the threshold in the negative direction to complete the feedback correction.

[0090] See appendix Figure 3 The present invention also proposes an adaptive optimization system for filling ratio based on parameter correlation neural network, comprising the following modules: The signal synchronization module is used to acquire the rheological shear force signal and micro-vibration acoustic wave signal continuously collected by the detection device, and perform signal alignment based on time delay compensation to generate a synchronization physical signal; The phase feature extraction module is used to analyze the high-frequency envelope features of the synchronous physical signal and extract the transient phase derivative. The compilation sequence generation module is used to trigger node mapping rules based on the comparison results of transient phase derivative and interference phase transition threshold, and generate node connection index sequence; The gated matrix construction module is used to parse the node connection index sequence to reconstruct the material node association and generate a dynamic topological adjacency matrix; The network instantiation module is used to inject the computational path of the graph node reconstruction using the dynamic topological adjacency matrix, instantiate the initial graph neural network, and generate the instance graph neural network. The forward inference module is used to call the instance graph neural network to perform forward graph convolutional inference calculations, decode the output intensity prediction residual value and the comprehensive cost evaluation vector; The iterative control module is used to combine the intensity prediction residual value and the comprehensive cost evaluation vector to perform multi-level target penalty optimization and output the target filling ratio instruction.

[0091] Each of the modules can be implemented in whole or in part through software, hardware, or a combination thereof. It supports hardware embedded in or independent of the processor in the computer device, and also supports software stored in the memory of the computer device, so that the processor can call and execute the operations corresponding to each of the above modules.

[0092] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for adaptive optimization of filling ratio based on parameter correlation neural network, characterized in that, Includes the following steps: S1. Acquire the rheological shear force signal and micro-vibration acoustic wave signal continuously collected by the detection device, and perform signal alignment based on time delay compensation to generate a synchronous physical signal; S2. Analyze the high-frequency envelope characteristics of the synchronous physical signal and extract the transient phase derivative; S3. Based on the comparison results of transient phase derivative and interference phase transition threshold, the node mapping rule is triggered to generate a node connection index sequence; S4. Parse the node connection index sequence to reconstruct the material node association and generate a dynamic topological adjacency matrix; S5. Inject the dynamic topological adjacency matrix to reconstruct the computation path of the graph nodes, instantiate the initial graph neural network, and generate the instance graph neural network; S6. Call the instance graph neural network to perform forward graph convolutional inference calculation, decode the output intensity prediction residual value and the comprehensive cost evaluation vector; S7. Combine the intensity prediction residual value and the comprehensive cost assessment vector to perform multi-level target penalty optimization and output the target filling ratio instruction.

2. The adaptive optimization method for filling ratio based on parameter correlation neural network according to claim 1, characterized in that, Step S1 includes the following steps: Extract the inter-channel time deviation between the rheological shear force signal and the micro-vibration acoustic wave signal; The cross-correlation integral algorithm is used to process rheological shear force signals and micro-vibration acoustic signals to quantify the time delay between channels. The rheological shear force signal is optimized by time axis translation and phase compensation using the time delay between channels, and then aligned with the micro-vibration acoustic wave signal to generate a synchronous physical signal.

3. The adaptive optimization method for filling ratio based on parameter correlation neural network according to claim 1, characterized in that, Step S2 includes the following steps: Transient intrinsic mode decomposition is performed on the data streams of each communication channel of the synchronous physical signal to remove the macroscopic steady-state low-frequency components and separate the high-frequency interference envelope with a frequency higher than the preset center frequency threshold. The analytical signals of the high-frequency interference envelope sequence corresponding to multi-component materials are calculated by Hilbert transform to generate dynamic cross-analytical signals and capture the extreme point information of the energy envelope of the signal. Based on the extreme point information, the first-order time difference value of the cross-phase function is obtained, and the transient phase derivative is calculated.

4. The adaptive optimization method for filling ratio based on parameter correlation neural network according to claim 1, characterized in that, Step S3 includes the following steps: Determine whether the transient phase derivative crosses the preset interference phase transition threshold, and convert the transient phase derivative that has passed the threshold screening into a dynamic step size factor; Based on the dynamic step size factor, the parameter node state in the preset one-dimensional node state array is updated, and the target node whose state identifier is updated to the active state is obtained. Collect the array index features of the target nodes and convert them into node index offsets. Then, concatenate the continuously generated node index offsets to construct a node connection index sequence.

5. The adaptive optimization method for filling ratio based on parameter correlation neural network according to claim 1, characterized in that, Step S4 includes the following steps: Parse and extract the consecutive index instructions contained in the node connection index sequence; The continuous indexing instructions are projected onto the preset two-dimensional node association space that represents the distribution of material types, and the edge connection weights between material nodes are initially determined. Remove invalid coupled lines in the two-dimensional node association space that do not match the coordinate position corresponding to the index command, and lock and retain valid association edges; Based on effective associated edges and preset active material nodes, a node connectivity graph is constructed, and a dynamic topological adjacency matrix is ​​encapsulated to adapt to transient mixed features.

6. The adaptive optimization method for filling ratio based on parameter correlation neural network according to claim 1, characterized in that, Step S5 includes the following steps: Initialize and create an initial graph neural network containing multiple graph nodes; The dynamic topological adjacency matrix is ​​extracted as a shape constraint that defines the dimension of the convolution tensor of the underlying graph; Extract shape constraints to update the empty connection architecture of the initial graph neural network; Based on shape constraints, feature computation channels are assigned to graph node pairs that establish connectivity, and feature propagation masking is performed on graph node pairs that do not have connectivity, thus generating instance graph neural networks.

7. The adaptive optimization method for filling ratio based on parameter correlation neural network according to claim 1, characterized in that, Step S6 includes the following steps: Obtain the standard material ratio baseline value and candidate adjustment range variables required for the target processing cycle; The standard material ratio baseline value and candidate adjustment range variables are fused and transmitted to the input layer of the instance graph neural network; The forward graph convolution extraction and activation function propagation are performed sequentially along the node connection channels of the instance graph neural network, and a preset validation set distribution threshold is applied to suppress feature value overflow. The multidimensional convergence tensor generated by the output layer of the aggregated instance graph neural network is used for feature dimensionality reduction, separation, decoding, and output of intensity prediction residuals and comprehensive cost evaluation vectors.

8. The adaptive optimization method for filling ratio based on parameter correlation neural network according to claim 1, characterized in that, Step S7 includes the following steps: Different convergence weight coefficients are preset for the intensity prediction residual value and the comprehensive cost assessment vector, and a fitness assessment function containing a cost penalty term is synthesized. Configure joint constraint boundary conditions to maximize the static pass rate of the filling body and minimize the cost of filling material per unit volume; The nonlinear heuristic algorithm is activated to calculate the fitness evaluation function in real time within the preset ratio adjustment range space, and to search for the target solution set containing the target adjustment range variable; Numerical target parameters are extracted from the target solution set, jointly calculated, and converted into target filling ratio instructions that conform to the communication protocol, which are then sent to the field control layer.

9. The adaptive optimization method for filling ratio based on parameter correlation neural network according to claim 2, characterized in that, Before performing cross-correlation integration on the rheological shear force signal and the micro-vibration acoustic signal, the following steps are also included: The rheological shear force signal and the micro-vibration acoustic signal are subjected to noise reduction preprocessing using a preset mean filtering algorithm to eliminate background noise interference. The absolute physical units carried by the data stream after noise reduction and preprocessing are removed by the zero-mean normalization equation, and a dimensionless scalar function matrix is ​​generated. The dimensionless scalar function matrix is ​​used as the recombined rheological shear force signal and the microseismic acoustic signal to perform cross-correlation integration, thus shielding cross-dimensional calculation errors.

10. The adaptive optimization method for filling ratio based on parameter correlation neural network according to claim 8, characterized in that, After extracting numerical target parameters from the target solution set, jointly calculating them, and converting them into target filling ratio instructions conforming to the communication protocol, which are then sent to the field control layer, the process also includes: Simultaneously initiate the data retrieval task to receive the actual strength measurement parameters obtained from the standard filling test blocks formed by on-site physical execution after standard curing; Calculate the numerical deviation term between the actual intensity measurement parameters and the intensity value predicted by the network model in the current execution cycle; The actual intensity measurement parameters are used as supervision labels and substituted into the loss function to calculate the gradient, driving the online incremental training of the weights of the bottom convolutional layers of the instance graph neural network. At the same time, the numerical deviation term is extracted as an error compensation correction factor to adjust the interference phase transition threshold.

Images