Filling cost decision method based on neural network multi-objective optimization

Abstract

The present application belongs to the technical field of filling cost decision, and relates to a filling cost decision method based on neural network multi-objective optimization, comprising the following steps: obtaining filling operation time sequence data and economic dynamic instruction fusion to generate basic data sequence; analyzing material cost flow calculation evolution evaluation quantity to generate economic optimization gradient characteristic flow; extracting compression strength test value for offset control to generate fatigue shrinkage penalty variable; injecting physical state parameter into physical model to introduce penalty variable to calculate safety boundary threshold value; deriving candidate proportioning matrix through pre-training neural network, combining economic characteristic flow and original detection correction quantity to determine joint optimization gradient, and generating controlled update result according to threshold value screening; and converting into material proportioning quota and issuing controlled equipment when reaching steady state. The present application solves the problems of multi-source real-time data synchronization, economic feedback fusion and physical boundary self-adaptive adjustment difficulty, and the problem that filling cost optimization and long-term safety are difficult to be considered simultaneously.

Description

Technical Field

[0001] This invention belongs to the technical field of filling cost decision-making, and relates to a filling cost decision-making method based on neural network multi-objective optimization. Background Technology

[0002] Backfilling operations are a crucial part of underground mining, primarily utilizing tailings, cementitious materials, and water to prepare backfill slurry, which is then transported to the underground goaf to control ground pressure. Since cementitious material consumption accounts for a significant portion of the total backfilling operating cost, achieving economic optimization of the backfill material ratio while meeting the physical and mechanical properties of the backfill and pipeline safety constraints is a critical technical challenge in mine production management.

[0003] Existing methods for determining fill mix proportions primarily rely on empirical formulas or static mix design schemes based on laboratory test data. Common implementation methods include technicians pre-setting fixed formula parameters such as the ash-sand ratio based on historical production data and design specifications, and then selecting these parameters according to the filling purpose; or using offline optimization software to input basic physical property parameters and then using mathematical models to calculate the minimum material usage to meet specific strength targets. These decision-making mechanisms have certain production applicability in scenarios with stable material sources and simple filling conditions.

[0004] The aforementioned static proportioning and offline calculation methods have certain technical limitations in practical applications. First, the fixed formulation model struggles to fully couple external economic signals such as raw material price fluctuations with real-time process conditions such as instantaneous differential pressure, making it difficult to automatically adjust material weights to mitigate costs in the face of drastic price changes. Second, traditional offline strength verification based on ideal conditions has a time lag, making it difficult for the decision-making system to perceive in real time the deterioration trend of backfill performance caused by tailings property fluctuations, increasing the probability of low strength or pipeline obstruction during long-term implementation. Furthermore, existing technologies typically treat physical safety boundary setting and cost optimization as separate processing steps, lacking a multi-objective collaborative mechanism, and are prone to deviating from dynamically changing pipeline safety limits when optimizing costs. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a filling cost decision-making method based on neural network multi-objective optimization.

[0006] The filling cost decision-making method based on neural network multi-objective optimization includes the following steps: S1, acquire the time series data and external economic dynamics instructions corresponding to the filling operation, and merge and splice the time series data and external economic dynamics instructions to generate a basic data sequence; S2, parse the historical material cost data stream contained in the basic data sequence to calculate the evolutionary evaluation quantity, and generate the economic optimization gradient feature stream based on the evolutionary evaluation quantity; S3, extract the compressive strength test values ​​from the basic data sequence feedback for offset comparison, and calculate and generate fatigue shrinkage penalty variables; S4, Load the physical state parameters contained in the basic data sequence into the preset physical model, introduce fatigue shrinkage penalty variables to perform boundary shrinkage calculation, and generate a safe boundary threshold. S5. The candidate allocation matrix is ​​derived through the control node based on the pre-trained neural network. The joint optimization gradient is determined by combining the economic optimization gradient feature flow and the original detection correction amount output by the neural network. The joint optimization gradient is then filtered according to the safety boundary threshold to generate a controlled update result. S6. When the controlled update result reaches a stable state, extract the steady-state variable group and convert it into a material proportion quota. Generate a proportion execution instruction based on the material proportion quota and send it to the controlled equipment.

[0007] A further aspect of the present invention, step S1, includes the following steps: Periodically read tailings concentration characteristic data, pipeline filling differential pressure data, and electromagnetic flow data to form time series data; By invoking material price change signals and batch planned consumption, external economic dynamics instructions are formed. External economic dynamics commands are upsampled according to a preset uniform sampling frequency to align with the time axis of the time series data; The upsampled external economic dynamics instructions are concatenated with time series data to generate a basic data sequence.

[0008] A further aspect of the present invention, step S2, includes the following steps: Extract material price sampling points within a time window from the basic data sequence to construct a historical data stream of material costs; The price change acceleration is obtained by performing second-order difference calculation on the historical data stream of material costs. The price change acceleration is converted into a penalty coefficient matrix using a nonlinear mapping function; Multiplying the penalty coefficient matrix by the baseline gradient vector yields the economic optimization gradient feature flow.

[0009] A further aspect of the present invention, step S3, includes the following steps: Receive compressive strength test values ​​with timestamps; By using timestamps, the flow rate decay value corresponding to the production period can be traced back in the basic data sequence; The comprehensive drift measure level is calculated based on the strength deviation of the compressive strength test value relative to the design threshold and the flow velocity decay value. The fatigue shrinkage penalty variable is calculated and generated based on the comprehensive drift level and the preset tolerance.

[0010] A further aspect of the present invention, step S4, includes the following steps: Input the current slurry ratio information and friction loss parameters from the basic data sequence into the physical model; The initial boundary is a pre-defined fluid dynamics boundary. The initial boundary is adjusted to obtain the dynamic shrinkage boundary; Based on the dynamic contraction boundary, the critical yield stress and ultimate plastic viscosity are output in the physical model to form a safety boundary threshold.

[0011] A further aspect of the present invention, step S5, includes the following steps: Generate a candidate ratio matrix based on the current ratio parameters; The joint optimization gradient is obtained by weighted summation of the original detection correction amount and the economic optimization gradient feature flow. Predict the physical parameters of each candidate scheme in the candidate ratio matrix; Determine whether each predicted physical parameter exceeds the safety boundary threshold; A usable mask is generated for the candidate solutions that have not exceeded the limit, and the mask is multiplied with the joint optimization gradient to derive the controlled update processing result.

[0012] A further aspect of the present invention, step S6, includes the following steps: Calculate the parameter update amount of the controlled update result in successive iterations; When the parameter update amount is lower than the preset convergence threshold, it is confirmed that a stable state has been reached, and the steady-state variable set is extracted. Extract the required material proportion quotas for each material from the steady-state variable set; According to the equipment interaction protocol, the material proportioning quota is converted into proportioning execution instructions and sent to the controlled equipment through the communication bus.

[0013] A further aspect of the present invention includes, before aligning the time axis of the time series data: Identify and read high-frequency noise in pipeline filling differential pressure data; Wavelet transform is applied to filter and remove wavelet coefficients corresponding to high-frequency noise; The wavelet coefficients are reconstructed to obtain the denoised filling differential pressure data.

[0014] A further aspect of the present invention involves calculating the overall drift measure level, including: Calculate the ratio of the difference between the test compressive strength value and the design target strength; Calculate the portion of the absolute value of the flow velocity decay value that exceeds the absolute value of the reference flow velocity decay rate, and determine the percentage difference accordingly. The weighted sum of the two difference ratios yields the overall drift measure.

[0015] A further aspect of the present invention involves determining whether each predicted physical parameter exceeds a safety boundary threshold, including: The predicted yield stress and plastic viscosity of each candidate scheme are compared with the corresponding upper and lower limits in the safety boundary threshold. If any predicted value exceeds the corresponding upper or lower limit range, the mask of the candidate solution is set to zero.

[0016] In summary, the present invention has the following beneficial technical effects: 1. This invention generates a basic data sequence containing multidimensional features by performing multidimensional noise reduction and frequency alignment transformation on physical time series and economic change signals. This technique achieves the structured fusion of real-time physical state parameters of the filling process and macroeconomic information, providing a unified data input basis for decision-making models. It helps the decision-making system to synchronously perceive the dynamic changes of the external environment and process status, and improves the lag problem caused by data fragmentation in traditional static proportioning decisions.

[0017] 2. This invention calculates a comprehensive drift measure by matching and cross-referencing time-lag-dependent compressive strength test values ​​with flow rate attenuation values ​​during the corresponding production period using timestamps. This processing method transforms a single quality inspection index into a quantitative evaluation quantity that takes into account fluid process deviations. This enables the system to identify potential deterioration trends in the performance of the filling material, and thereby generate a fatigue shrinkage penalty variable, providing a quantitative basis for tightening physical control boundaries and reducing the risk of long-term quality instability.

[0018] 3. This invention introduces fatigue shrinkage penalty variables into the fluid dynamics physical model for bidirectional boundary calculation. Combined with the native detection correction quantity output by the pre-trained neural network, the safety boundary thresholds, such as the critical yield stress, can be dynamically scaled according to the degree of system degradation. Simultaneously, by converting the generated safety boundary thresholds into physical intervention masks, candidate ratio paths that predict out-of-bounds errors are blocked, ensuring that the parameter iterative optimization process remains within a feasible physical operating space, thus improving the safety of the final optimized ratio in actual pipeline transportation.

[0019] 4. This invention constructs a joint optimization correction mechanism that integrates the economic optimization gradient feature flow and the dynamic safety boundary threshold, simultaneously applying dual constraints of market price mutations and long-term quality drift during the iterative update of model parameters. This mechanism ensures that the adjustment direction of the filling ratio is guided by both the material cost reduction gradient and the hard mask constraint of the safety boundary, forming data synergy between seeking material economic optimization and maintaining dynamic process safety, thus alleviating the problem of solely pursuing cost reduction while ignoring dynamic physical red lines. Attached Figure Description

[0020] 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.

[0021] Figure 1 A flowchart illustrating an embodiment of this application is disclosed.

[0022] Figure 2 The mapping relationship curve between price change acceleration and penalty coefficient in the embodiments of this application is disclosed.

[0023] Figure 3 The dynamic shrinkage trend diagram of the critical yield stress boundary as a function of fatigue shrinkage penalty variable in the embodiments of this application is disclosed.

[0024] Figure 4 Structural schematic diagrams of embodiments of this application are disclosed. Detailed Implementation

[0025] The following is in conjunction with the appendix Figures 1-4 A preferred description of the present invention is provided below.

[0026] See attached document Figure 1 This invention proposes a filling cost decision-making method based on neural network multi-objective optimization, comprising the following steps: S1, acquire the time series data and external economic dynamics instructions corresponding to the filling operation, and merge and splice the time series data and external economic dynamics instructions to generate a basic data sequence; S2, parse the historical material cost data stream contained in the basic data sequence to calculate the evolutionary evaluation quantity, and generate the economic optimization gradient feature stream based on the evolutionary evaluation quantity; S3, extract the compressive strength test values ​​from the basic data sequence feedback for offset comparison, and calculate and generate fatigue shrinkage penalty variables; S4, Load the physical state parameters contained in the basic data sequence into the preset physical model, introduce fatigue shrinkage penalty variables to perform boundary shrinkage calculation, and generate a safe boundary threshold. S5. The candidate allocation matrix is ​​derived through the control node based on the pre-trained neural network. The joint optimization gradient is determined by combining the economic optimization gradient feature flow and the original detection correction amount output by the neural network. The joint optimization gradient is then filtered according to the safety boundary threshold to generate a controlled update result. S6. When the controlled update result reaches a stable state, extract the steady-state variable group and convert it into a material proportion quota. Generate a proportion execution instruction based on the material proportion quota and send it to the controlled equipment.

[0027] In one embodiment of the present invention, step S1 includes the following steps: Periodically read tailings concentration characteristic data, pipeline filling differential pressure data, and electromagnetic flow data to form time series data; By invoking material price change signals and batch planned consumption, external economic dynamics instructions are formed. External economic dynamics commands are upsampled according to a preset uniform sampling frequency to align with the time axis of the time series data; The upsampled external economic dynamics instructions are concatenated with time series data to generate a basic data sequence.

[0028] Step S1 in this embodiment is executed on the data processing server deployed at the mine backfilling monitoring center. The data processing server establishes a communication connection with the field programmable logic controller (PLC) network via an industrial Ethernet interface, and accesses and captures a set of time-series data in real time; the process is as follows: Data is periodically read from multiple sensor ports connected to the PLC. These sensors include an ultrasonic concentration meter installed on the outlet pipe of the filling preparation mixing tank to obtain the current tailings concentration characteristics. Differential pressure transmitters installed on the pipeline between the inlet and outlet of the filling industrial pump are used to obtain the differential pressure during pipeline filling; and electromagnetic flowmeters installed on the main delivery pipeline are used to obtain electromagnetic flow records. The analog signals collected by these sensors, such as 4-20mA current signals, are converted into digital quantities by the PLC's built-in analog-to-digital converter, and packaged into data frames at a preset physical data sampling frequency and sent to the data processing server to form a set of original physical quantity time series.

[0029] At the same time, the data processing server executes parallel scheduling tasks, synchronously calling the material information module in the enterprise resource planning system through the application programming interface to obtain price change signals of materials such as cement and fly ash covering the target optimized production cycle.

[0030] At the same time, the server also accesses the manufacturing execution system to retrieve pre-prepared batch plans for the filling of specific goaf areas, which together constitute external economic dynamics instructions.

[0031] After acquiring time-series data and external economic dynamics instructions, the data processing server initiates the data preprocessing pipeline. In this pipeline, multi-dimensional noise reduction and filtering are first performed on the time-series data. For example, wavelet transform is applied to the highly volatile differential pressure signal of pipeline filling. By decomposing the signal into different scales, thresholds are set for the wavelet coefficients corresponding to high-frequency noise, and they are either zeroed out or compressed for filtering. Then, the signal is reconstructed to filter out noise introduced by pumping pulsations or measurement interference.

[0032] For external economic dynamics, a moving average filter is applied to smooth the data, eliminating short-term trading noise in price data and revealing its long-term trend.

[0033] Subsequently, the information sampling frequency is uniformly aligned and transformed. Considering that the sampling frequency of time series data is much higher than the update frequency of external economic dynamics instructions, the processing module adopts a forward filling strategy to upsample low-frequency economic data, such as daily updated price values, and fill them into each physical data sampling time point.

[0034] At each sampling time point, the denoised real-world physical dimension parameters and the aligned cost dimension state information are horizontally concatenated in a predetermined order to form a high-dimensional feature vector. The high-dimensional feature vectors from all time points are then vertically stacked in temporal order to reconstruct and generate a basic data sequence that integrates physical state parameters and economic state information.

[0035] Time series data is a vector set describing the physical state of the filling slurry during pipeline transportation. Specifically, it includes tailings concentration characteristics (in kPa, typically ranging from 65% to 78% by mass); pipeline filling differential pressure (in kPa, reflecting friction loss during slurry transport); and electromagnetic flow records (in kPa). This characterizes the volume of slurry transported per unit time.

[0036] External economic dynamics commands are a set of parameters describing the economic boundary conditions related to filling operations, specifically including material price change signals, and their units are... ; and the batch planned consumption, in units of The physical data sampling frequency is set based on a trade-off between the real-time requirements of the filling system control and the cost of data storage. A typical setting is 1Hz, which is sufficient to capture macroscopic changes in the slurry flow pattern without excessively increasing the computational burden.

[0037] The threshold for wavelet transform is adaptively set based on the noise level of the signal. For example, a general threshold strategy based on signal standard deviation estimation can be used to effectively filter out noise while preserving the main features of the signal. The window size of the moving average filter used to smooth the price signal is set according to the price update cycle. Assuming that the price data is updated daily, the window size can be set to 7, that is, using the 7-day average price to reflect the short-term price trend.

[0038] Basic data sequence yes The matrix, where The length of the time series. The total dimension of the merged features can be expressed mathematically as follows:

[0039] Among them, at any sampling time row vectors for:

[0040] in, , , They are respectively The tailings concentration, pipeline filling differential pressure, and electromagnetic flow rate after noise reduction processing are recorded. and These are the smoothed prices of cement and fly ash after alignment at that moment, respectively. This represents the batch planned consumption amount after alignment at the corresponding time.

[0041] In one example of the present invention, within a continuous 3 seconds of a filling operation, the data processing server obtains the raw physical quantity time series data from the PLC network as follows: At that time, the tailings concentration was 72.0%, the differential pressure was 850 kPa, and the flow rate was... ; Due to the pump's pulsation, the differential pressure spikes at 875 kPa, while other values ​​remain largely unchanged. At any given time, all parameters recover to near The level of real-time data. Meanwhile, the server retrieves the daily cement price from the ERP system. The price of fly ash is The planned filling volume for the current batch is obtained from the MES system. The server first processes the differential pressure time series. Apply wavelet denoising to The noise level at any given time was corrected from 875 kPa to a more reasonable 852 kPa. Next, frequency alignment was performed, adjusting the daily cement price... fly ash price and total batch volume Fill them into At these three points in time, within a few seconds, the server ultimately pieces together the data to generate the initial three rows of the basic data sequence. The row vector corresponding to time step is ; The row vector corresponding to time step is ; The row vector corresponding to time step is These three row vectors form part of the basic data sequence, providing a structured input that integrates real-time physical state and macroeconomic context for subsequent steps.

[0042] In one embodiment of the present invention, step S2 includes the following steps: Extract material price sampling points within a time window from the basic data sequence to construct a historical data stream of material costs; The price change acceleration is obtained by performing second-order difference calculation on the historical data stream of material costs. The price change acceleration is converted into a penalty coefficient matrix using a nonlinear mapping function; Multiplying the penalty coefficient matrix by the baseline gradient vector yields the economic optimization gradient feature flow.

[0043] After receiving the basic data sequence generated in step S1, the data processing server parses the economic information contained therein and generates gradients to guide subsequent optimization.

[0044] The server performs a slicing operation, selectively extracting columns related to price variables from the underlying data sequence matrix. These columns constitute a historical data stream of material costs, such as time series data containing cement and fly ash prices.

[0045] For each material price time series, since the underlying data series has already undergone high-frequency forward filling, the server uses a step-wise extraction method based on the original economic data update cycle to form a time window, obtaining the actual macroeconomic change nodes as valid data points, and performing second-order differencing calculations on the data within this window. This estimation is achieved by calculating the discrete second-order central difference of the price series, thereby quantifying an evolutionary assessment quantity, specifically the acceleration of price changes, i.e., evolutionary potential energy.

[0046] Based on the price change acceleration calculated in the previous step, the data processing server infers the evolution trend of the unit price of each material and its abrupt change points. (See also...) Figure 2 The mapping relationship between price change acceleration and penalty coefficient is shown. When the price change acceleration of a certain material is positive and its absolute value exceeds the preset intensification threshold, the system determines that the price of the material is at a sudden intensification point of accelerated rise.

[0047] Conversely, if the value is negative and its absolute value exceeds the intensification threshold, it is determined to be an accelerated decline. Based on this judgment, the system maps and assigns values ​​to the price change acceleration through a preset nonlinear mapping function, generating a corresponding price intervention dimension penalty coefficient for each independent material delivery channel. All these coefficients together constitute a penalty coefficient matrix. The purpose of this matrix is ​​to transform the abstract price change trend into quantitative weights usable in subsequent optimization models.

[0048] The system loads this penalty coefficient matrix and uses it as a weight vector to adjust the weights of different material costs during optimization. Specifically, this matrix is ​​multiplied element-wise with the baseline optimization gradient vector to modulate the output economic optimization gradient feature stream. This feature stream encodes the strength of the optimization tendency towards different material costs under the current economic environment and is retained by the system to activate the intervention state indication command in subsequent steps.

[0049] In this step, the acceleration of price changes Regarding the first Such materials at all times price The calculation formula is as follows:

[0050] in This is the sampling time interval. In the calculation, if... Since it is a constant, it can be normalized. Penalty coefficient matrix. Each penalty coefficient in Depend on Through mapping function get:

[0051] Typical mapping function It can be a scaled and translated Sigmoid function:

[0052] Economic Optimization Gradient Feature Flow By using the penalty coefficient matrix With reference gradient vector Generate by performing Hadamard product operation:

[0053] Among them, the price change acceleration is a quantitative value that characterizes the acceleration of the trend of material price change. Its positive or negative sign indicates the concavity or convexity of the price curve, that is, whether it is accelerating upward or accelerating downward.

[0054] The acceleration threshold criterion is a threshold criterion for the acceleration of price changes. The system is based on statistical analysis of daily material price data from at least one year of historical data, using the 95th percentile of the second difference to identify drastic price fluctuations that exceed normal ranges. (Penalty coefficient matrix) It is a row vector with dimensions equal to the number of bulk material types being monitored, where each element is a dimensionless penalty coefficient. Its range of values ​​is determined by the mapping function parameters. and Control, among which To ensure that the exponent term is a dimensionless pure number, an adjustment coefficient with a corresponding inverse vector dimension is set, for example, by setting... , , It refers to the base time unit for updating economic data, such as 1 day. It can amplify the penalty coefficient when prices rise rapidly, while keeping the coefficient close to the base value when prices fall.

[0055] For example, for monitoring both cement and fly ash, this matrix would be... Economic optimization gradient characteristic flow It is the final output vector carrying information about the direction of economic optimization, which consists of the penalty coefficient matrix and the benchmark gradient vector generated based on the normalization of the benchmark prices of each material. Multiply them to get the result. This reflects the absolute difference in material costs. For example, the benchmark cost of cement is much higher than that of fly ash, and its gradient component is larger. The physical meaning of this multiplication operation is to simultaneously transmit the absolute cost weight at the economic level and the pressure of dynamic price changes to the subsequent optimization algorithm.

[0056] In one example of the present invention, to follow the scenario of step S1, it is assumed that after the data processing server extracts the stride truncation, it extracts the corresponding three consecutive economic update cycles from the basic data sequence, such as the cement price sequence for three consecutive days. and fly ash price series Here, the sampling time interval is defined as the reference time unit. At the second sampling time, the server calculates the acceleration of price changes for both materials. For cement, For fly ash, Assuming a preset activation threshold Due to cement The value is 5, which is greater than The system determines that cement prices are in a state of sudden and intense fluctuation. Next, a mapping function is used. Calculate the penalty coefficient. The penalty coefficient for cement is... The penalty coefficient for fly ash is... Therefore, the generated penalty coefficient matrix is ​​as follows: Finally, a normalized benchmark gradient vector is calculated based on the absolute prices of cement and fly ash. Assuming the current absolute cost of cement accounts for 0.86 of the weight and fly ash 0.14, that is... Then the economic optimization gradient feature flow generated by the server modulation is The results were retained, indicating that subsequent optimization processes should favor reducing the cost of high-priced and rapidly increasing cement usage with a very high weight of 2.449, while the weight for reducing fly ash was only 0.28, taking into account both the cost base and the economic signals of market fluctuations.

[0057] In one embodiment of the present invention, step S3 includes the following steps: Receive compressive strength test values ​​with timestamps; By using timestamps, the flow rate decay value corresponding to the production period can be traced back in the basic data sequence; The comprehensive drift measure level is calculated based on the strength deviation of the compressive strength test value relative to the design threshold and the flow velocity decay value. The fatigue shrinkage penalty variable is calculated and generated based on the comprehensive drift level and the preset tolerance.

[0058] As the system continues with step S3, the data processing server performs in-depth analysis of the basic data sequence generated in the preceding steps to assess the system's long-term performance drift. First, the server initiates a feature separation and extraction task, scanning the basic data sequence and identifying characteristic parameter signals containing hysteresis. Specifically, the system polls an interface dedicated to receiving hysteresis quality testing data, which periodically receives compressive strength test values ​​uploaded from the mining laboratory. These sampled values, such as the uniaxial compressive strength test results of infill samples after reaching 28 days of standard curing conditions, inherently possess the characteristic of being in a state of hysteresis. Each sampled value is accompanied by its original production batch timestamp metadata, which the server uses for targeted retrieval.

[0059] After obtaining the sampled value, the server performs a reverse backtracking timeline alignment and correlation operation. This involves precisely locating the historical data segment corresponding to the entity sample at the production moment within the complete basic data sequence based on the timestamp metadata. Within this historical data segment, the server further extracts and encapsulates the flow rate attenuation value recorded within the flow path of the basic data sequence. Then, through comprehensive correlation and cross-validation, the server compares and analyzes the delayed compressive strength test value with the flow rate attenuation value during real-time production to determine the quantitative indicator, namely the comprehensive drift metric, which measures the degree of potential performance degradation.

[0060] The server compares the calculated overall drift measure with the preset tolerance set in the system configuration, and calculates the difference in deviation from the tolerance. This parameter, as the final result, is output by the system to form a fatigue shrinkage penalty variable, providing a basis for the dynamic adjustment of the physical boundary in subsequent steps.

[0061] In this step, the overall drift measurement level is determined. It is a dimensionless index that integrates both quality and process deviations, and its calculation formula is:

[0062] Subsequently, the difference exceeding the tolerance limit From the overall drift measurement level With preset tolerance The calculation shows that:

[0063] Fatigue contraction penalty variable Through the After scaling, we get:

[0064] In the formula, the compressive strength test value This is the compressive strength value obtained by testing the filling sample according to the relevant national standards, with the unit being MPa; the design target compressive strength value. The unit is MPa; flow rate attenuation value This value is obtained by performing a first-order difference operation on the time series of electromagnetic flow records during the production of this sample and calculating the mean value. It reflects the rate of flow velocity decay in the pipeline due to settling or premature hydration of the slurry, and the unit is 1. A large negative value indicates that the flow velocity decreases too quickly; the reference flow velocity decay rate The unit is Weighting coefficients and This is a preset dimensionless constant, and the sum of the two is 1. It is used to balance the importance of quality deviation and process deviation in the comprehensive evaluation. For example, it can be set to... and ; Overall drift measurement level The dimensionless scalar value obtained through normalization; preset tolerance. This is a dimensionless threshold value, set based on the minimum requirements for backfill performance in mine safety regulations and the statistical tolerance range for process stability; for example, it can be set to 0.15. The difference exceeding the tolerance range... A dimensionless scalar value; scaling factor It is a constant greater than 1, used to amplify the penalty force of deviation on subsequent control, for example, set to 1.5; fatigue contraction penalty variable. It is a dimensionless scalar value.

[0065] In one example of the present invention, it is assumed that the system receives a compressive strength test value at 28 days of age. Its corresponding design target strength is The server traces back the basic data sequence based on the production timestamp of the sample, extracts the flow data during the production period, and calculates the flow rate decay value. The system's preset reference flow rate attenuation rate is... The server first calculates the overall drift metric and substitutes it into the weights. and The calculation result is This value indicates that the system's overall drift has reached 51.3%. The server then compares this value with the preset tolerance. Compare the values. Calculate the difference exceeding the tolerance. Finally, use the scaling factor. The fatigue shrinkage penalty variable was calculated to be: This dimensionless value will be output to the next step to dynamically tighten the boundaries of physical constraints in response to identified trends of system performance degradation.

[0066] In one embodiment of the present invention, step S4 includes the following steps: Input the current slurry ratio information and friction loss parameters from the basic data sequence into the physical model; The initial boundary is a pre-defined fluid dynamics boundary. The initial boundary is adjusted to obtain the dynamic shrinkage boundary; Based on the dynamic contraction boundary, the critical yield stress and ultimate plastic viscosity are output in the physical model to form a safety boundary threshold.

[0067] First, the server extracts the latest current slurry mix information from the basic data sequence, including the real-time mass percentages of cement, fly ash, tailings, and water in the slurry. Simultaneously, it uses the pipeline filling differential pressure and electromagnetic flow records from the basic data sequence, combined with pre-stored pipeline geometric parameters such as pipe diameter and length, to calculate the friction loss parameters in real time. These two sets of data are then integrated and injected as input into a physical model embedded with fixed fluid dynamics mechanisms. This kernel structure is not a black-box model trained with data, but rather a white-box equation that solidifies the pipeline transport characteristics of non-Newtonian fluids, such as the Darcy-Weisbach head loss equation based on Bingham fluid constitutive relations, and the hydration mechanism equation describing the hydration reaction kinetics of cementitious materials.

[0068] The data processing server guides and accepts the fatigue shrinkage penalty variable calculated and delivered in step S3. This variable, as a constraint scaling factor, is applied directly to the boundary parameter calculation layer of the physical model in the form of a two-way multiplicative coefficient.

[0069] See Figure 3The fatigue shrinkage penalty variable, as shown, affects the bidirectional boundary shrinkage of the critical yield stress. Specifically, this variable is used to perform bidirectional boundary shrinkage calculations on a preset initial feasible region. For example, the system presets an initial safe upper and lower limit for the critical yield stress. The safe upper limit is used to prevent pipeline blockage, and the safe lower limit is used to prevent slurry segregation and insufficient strength. The fatigue shrinkage penalty variable will directly and dynamically reduce the upper limit while dynamically increasing the lower limit, supplemented by a lower limit floor to prevent excessive shrinkage lock-up, forming a more stringent dynamic safe feasible region that adapts to the potential deterioration trend of the current system. Under this shrinkage boundary constraint, the physical model, combined with the currently input slurry ratio information, dynamically calculates and outputs the upper and lower limits of the critical yield stress after shrinkage deformation, as well as the limiting plastic viscosity to resist shear and stick-slip of the pipeline inner wall, to constitute the safe boundary threshold.

[0070] Finally, these calculated scalar clusters of interception filters, which are not allowed to be exceeded in subsequent optimization explorations, are uniformly archived for subsequent steps to define the boundaries of the decision space.

[0071] In this step, the fatigue contraction penalty variable is applied. Adjusted dynamic fluid dynamics boundary parameters, such as the adjusted upper limit of critical yield stress. and lower limit It is calculated using the following formula:

[0072]

[0073] The Bingham fluid model used in the core structure describes the shear stress. With shear rate Relationship:

[0074] In the formula, The preset safety dead zone margin is used to forcefully ensure that the adjusted upper limit is always greater than the lower limit, thus avoiding deadlock due to the collapse of the optimization exploration space. This represents the fatigue contraction penalty variable, which is a dimensionless quantity. and These represent the upper and lower limits of the safety limit for the reference critical yield stress, respectively, both in Pa. and This represents the preset ratio of the maximum contraction to the maximum rise baseline. It is a dimensionless quantity used to prevent deadlock caused by excessive boundary contraction. and This represents the adjusted upper and lower limits of the critical yield stress, both in Pa. This represents shear stress, with units of Pa. This represents the actual yield stress predicted based on the current slurry mix ratio, in Pa. This indicates the predicted actual plastic viscosity based on the current slurry ratio, in Pa. s; Represents shear rate, in units of The safety boundary threshold is the set of critical parameters obtained through bidirectional adjustment.

[0075] It should be noted that the current slurry ratio information refers to the real-time mass fraction vector of each component of the slurry extracted from the basic data sequence, for example... Friction loss parameters are calculated in real-time from the pressure difference and flow rate data in the basic data sequence using fluid dynamics formulas, yielding the Reynolds number or friction coefficient. The physical model is a computational engine integrating well-known physical and chemical models; its parameters are fixed and it does not participate in the iterative training of machine learning.

[0076] Initial boundaries are a set of static safety thresholds set during the system design phase based on extensive experimental data and safety specifications, such as the maximum allowable yield stress. The minimum allowable yield stress is 50 Pa. The safety boundary threshold is 15 Pa. This is the final output vector of this step, which contains the dynamic safety upper and lower limits of all key physical properties under the current operating conditions, for example... These values ​​will be passed to subsequent optimization algorithms as hard constraints.

[0077] In one example of the present invention, continuing with the foregoing embodiments, the data processing server receives the fatigue shrinkage penalty variable calculated in step S3. The server reads the current slurry mix ratio from the basic data sequence as: 15% cement, 20% fly ash, 35% tailings, and 30% water. The maximum allowable yield stress is set in the initial boundary parameters defined by the system. Minimum allowable yield stress The server first performs bidirectional shrinkage calculations on these baseline boundaries using fatigue shrinkage penalty variables. Assume the system's preset scaling limit is... , The adjusted upper limit of the critical yield stress is: The adjusted lower limit of the critical yield stress was originally calculated to be 23.167 Pa. To prevent this lower limit from exceeding the upper limit and triggering deadlock, the system introduces a preset [measurement / mechanism]. The lower limit is calculated using a safety net method, and the actual adjusted lower limit is [value missing]. Subsequently, the physical model predicts the actual yield stress based on the current composition information. The safety boundary threshold generated by the system is the conflict-free vector that includes these adjusted boundaries. In subsequent steps, dynamically tightened upper and lower limits will intercept divergent candidate formulations.

[0078] In one embodiment of the present invention, step S5 includes the following steps: Generate a candidate ratio matrix based on the current ratio parameters; The joint optimization gradient is obtained by weighted summation of the original detection correction amount and the economic optimization gradient feature flow. Predict the physical parameters of each candidate scheme in the candidate ratio matrix; Determine whether each predicted physical parameter exceeds the safety boundary threshold; A usable mask is generated for the candidate solutions that have not exceeded the limit, and the mask is multiplied with the joint optimization gradient to derive the controlled update processing result.

[0079] First, the server activates and starts the control nodes configured within the overall computing control architecture, which are based on pre-trained neural networks, such as multilayer perceptrons or deep reinforcement learning networks. In this embodiment, the pre-trained neural network uses a multilayer perceptron architecture to construct a physical performance proxy model. Its network structure includes an input layer, at least two fully connected hidden layers, and an output layer. The input layer nodes receive the current state variables, specifically the current filling material ratio vector, such as the mass fractions of cement, fly ash, tailings, and water. The output layer nodes output the predicted physical performance indicators, such as predicted yield stress and predicted compressive strength.

[0080] The pre-trained neural network is trained offline using a historical filling operation dataset derived from actual batch ratio records and corresponding laboratory physical performance test results. During training, mean squared error is used as the loss function, and the Adam optimizer updates the neuron weights until the model's prediction error converges. In the current online decision-making phase, the neural network's weights are frozen, serving solely as a function mapper for forward inference and backward gradient calculation.

[0081] Based on the pre-trained neural network described above, the system executes a downlink inference and reverse link correction process. The system inputs the current filling material ratio into the network for forward propagation to obtain the current predicted physical properties. Subsequently, the system constructs a target loss function. This function calculates the sum of squares of the deviations between the current predicted physical properties and the design target properties, such as the target yield stress and the target compressive strength.

[0082] To narrow this gap, the system calculates the native detection correction amount. Unlike traditional backpropagation for updating network weights, this system utilizes the chain rule of backpropagation to keep the network weights fixed while calculating the target loss function. The gradient of the partial derivative with respect to the input layer variables, i.e., the current material proportion vector:

[0083] in, This is the current material proportion vector. This indicates the direction of ratio adjustment that can most quickly approach the target performance in terms of pure physical performance.

[0084] Based on this, the system uses the current material ratio as a reference point and follows the original detection correction amount. Discretize the indicated gradient direction and its surrounding feasible region by step-size sampling to derive a set of candidate proportion matrices with a divergent output structure. Each row of the matrix represents a potentially feasible next material proportion adjustment scheme.

[0085] Simultaneously, the system retrieves and loads the economic optimization gradient feature stream generated in step S2, uses this feature stream to construct an economic optimization guiding force, and superimposes it onto the original detection correction amount to form a joint optimization gradient that integrates performance optimization and cost orientation. After determining the joint adjustment direction, the system immediately activates the safety boundary threshold generated in step S4.

[0086] For each candidate material proportioning adjustment scheme mentioned above, the system predicts its physical state after execution and compares it with a safety boundary threshold. If the predicted physical state of a candidate scheme touches or exceeds the system's fault tolerance limit—for example, if the predicted yield stress exceeds the dynamically set upper limit or falls below the dynamically set lower limit—the system generates a physical intervention mask. This mask, through a multiplication operation, blocks or zeros out update paths leading to the infeasible scheme.

[0087] If the mask value of all candidate solutions is set to zero, meaning all paths with added economic optimization have crossed the physical red line, the system triggers a fallback mechanism, temporarily blocking the guiding force of economic optimization and reducing the economic impact coefficient. This forces the system to rely solely on the neural network's native probe corrections to retreat to a safe physical operating point, ultimately completing controlled update processing and ensuring that all actual adopted adjustments are within physical safety boundaries and avoiding iterative deadlock.

[0088] This step incorporates an economically oriented joint gradient. From native detection correction amount With economic optimization gradient feature flow Weighted summation yields:

[0089] For any candidate formulation Its corresponding physical intervention mask value The generation logic is as follows:

[0090] in, It is used to predict candidate formulation schemes. The function of the corresponding physical performance vector, and and This is the set of lower and upper limits parsed from the safety boundary threshold vector.

[0091] For any candidate solution, the matching variable The controlled update iteration process is represented as:

[0092] in, This is the physical intervention mask value (0 or 1) corresponding to the candidate solution. If this happens, the economic cost reduction update path leading to the solution is completely blocked, retaining only the original correction amount or directly eliminating the branch. After completing the above controlled update iteration, the updated matching variables... Each material component undergoes normalization to ensure that the sum of the proportions of each material is always equal to the specified value. .

[0093] The candidate matching matrix is A matrix of dimension, where It is the number of candidate solutions. It is an adjustable quantity of material composition. Native detection correction amount. It is an update direction vector calculated in the current iteration step by the built-in neural network model using optimization algorithms such as backpropagation, which only considers narrowing the gap between the physical performance index and the target value.

[0094] The economic optimization gradient characteristic flow establishes an economic optimization guiding force that forms a virtual potential field. Its physical meaning is to transform the downward trend of economic costs into an attractive field in the optimization space, guiding the update of candidate allocation variables towards lower costs. The economic influence coefficient λ is a preset hyperparameter, typically ranging from 0.05 to 0.3, used to adjust the weight of cost optimization in the overall objective.

[0095] The controlled update processing of physical intervention mask dot product is a computational mechanism that ensures that only updates whose predictive performance falls within the bidirectional shrinkage safety domain can be applied to the iterative adjustment of the matching state, thereby strictly adhering to the safety bottom line while pursuing economic goals.

[0096] In one example of the present invention, following the preceding steps, the data processing server has obtained the economic optimization gradient feature flow. This corresponds to the decreasing trend in the costs of cement and fly ash, respectively, and includes the adjusted lower limit of the critical yield stress. and upper limit Security boundary threshold Suppose the current optimization algorithm calculates the native detection correction amount aimed at improving the intensity as follows: The negative sign indicates a desire to increase the amount of cement and fly ash used. An economic impact coefficient is then set. The joint gradient is It is evident that the economic orientation successfully reversed the original instruction to increase cement usage to an instruction to decrease it. Simultaneously, because the sign of the fly ash gradient was not reversed, the trend of supplementing strength using inexpensive fly ash was preserved. Now, the decision node module is in the current mix proportion... Two candidate adjustment schemes were generated based on the data. Candidate Scheme 1: Reduce cement usage by 0.1 units and increase fly ash usage by 0.02 units. The system predicts this scheme will result in a slurry yield stress of 22.0 Pa. Candidate Scheme 2: To further reduce material costs, significantly reduce cement usage by 0.2 units and reduce fly ash usage by 0.05 units without constraint. The system predicts this scheme will result in a slurry yield stress of 20.0 Pa. Next, a safety boundary threshold was used for verification. For Candidate Scheme 1, its predicted yield stress of 22.0 Pa strictly falls within the specified range. Within the safe zone, therefore its mask value For candidate scheme 2, its predicted yield stress of 20.0 Pa falls below the lower limit of 21.775 Pa, therefore its mask value... Finally, during the execution of variable updates... At that time, the radical cost reduction path leading to candidate solution 2 will be blocked by the mask value 0, and only the update direction of candidate solution 1 will be retained. Thus, the system performs cost reduction operations under economic drive, while being intercepted by the upper and lower limits of the physical red line to avoid the extreme state that would lead to pipeline blockage or failure to meet the strength standard.

[0097] In one embodiment of the present invention, step S6 includes the following steps: Calculate the parameter update amount of the controlled update result in successive iterations; When the parameter update amount is lower than the preset convergence threshold, it is confirmed that a stable state has been reached, and the steady-state variable set is extracted. Extract the required material proportion quotas for each material from the steady-state variable set; According to the equipment interaction protocol, the material proportioning quota is converted into proportioning execution instructions and sent to the controlled equipment through the communication bus.

[0098] After completing the iterative update and decision correction in step S5, the data processing server enters the final execution instruction generation stage, i.e., step S6. In this stage, the server continuously tracks and evaluates the computational settling optimization process of the iterative model after undergoing controlled update processing using physical intervention mask dot multiplication. This process manifests as the system's inference engine traversing along the exploration path defined by the joint gradient, within a safe boundary that does not trigger warning states, and its trajectory position distribution gradually approaches a dynamic limit stable state.

[0099] The server monitors the rate of change of model parameter updates. When this rate of change falls below a preset convergence threshold for several consecutive iterations, the system determines that the computation has reached equilibrium and solidified, forming a steady-state variable set. After capturing this steady-state variable set, the server decodes it to extract the final determined material proportion quota. This material proportion quota represents the target filling material proportion scheme found by the system under all current economic, physical, and historical performance constraints.

[0100] Finally, to achieve direct control of the on-site equipment, the data processing server repackages and converts the extracted material proportion quotas according to pre-configured equipment interaction protocols compatible with the underlying associated electromechanical equipment. This conversion process transforms pure material proportion values, such as cement 350kg / m³, into... 3 This is converted into instruction code that can be recognized and executed by the underlying mechanical equipment. For example, it is converted into an instruction to control the output frequency of the cement screw feeder inverter to 35Hz and maintain it for 120 seconds. This instruction is ultimately sent to the PLC or distributed control system (DCS) at the mine site via an industrial fieldbus, such as the Modbus-TCP protocol, to complete the scheduling and control of specific feeding equipment and realize closed-loop automated filling operations.

[0101] The stable state of a controlled update result refers to the sequence of search points in the multidimensional parameter space by the iterative optimization algorithm. Ideally, this sequence will eventually converge to a local or global minimum of the objective function, which is the equilibrium final state. Convergence threshold. It is a small positive number used to determine whether the iteration has ended; for example, it can be set to 0. When the L2 norm of the parameter update vector is less than this value for multiple consecutive periods, it is considered convergent.

[0102] The steady-state variable set is the final set of values ​​for all parameters of the converged model, which fully defines the optimal decision. The material proportioning quota is a subset of parameters directly related to the material proportioning, typically expressed as the mass or volume of various solid materials per cubic meter of filling slurry, for example, {cement: 350 kg / m³}. 3 Tailings: 1400 kg / m³ 3 Water: 450L / m3}

[0103] The equipment interaction protocol is a pre-defined mapping rule base that defines in detail the conversion logic from abstract data quotas to specific equipment control parameters. For example, the rule base stores the calibration curve equations between the control parameters of each feeding device, such as frequency or valve opening and material flow rate. The proportioning execution instruction is a structured data message whose format strictly follows the communication protocol of the lower-level device. It includes information such as device address, function code, register address, and control values ​​to be written, ensuring that the instruction can be accurately parsed and executed.

[0104] In one example of the present invention, following the scenario of step S5, after multiple iterative calculations, the amount of parameter updates in the system gradually decreases. In the three most recent iteration cycles, the L2 norms of the parameter update vectors are respectively... , , All are less than the preset convergence threshold. The server determines that the model has converged and reached a final equilibrium state. At this point, the steady-state variable set is captured, and the server decodes the material proportions as follows: 325 kg of cement, 150 kg of fly ash, 1200 kg of tailings, and 525 kg of water are required per cubic meter of slurry. Next, the server queries its internal device interaction protocol. The protocol defines the flow rate of the cement screw feeder. With frequency converter The relationship is To meet the requirement of 325kg / m 3 And the total flow rate is 200 Based on the production requirements, the required cement feeding rate is calculated to be approximately 18.06 kg / s. Therefore, the required inverter frequency is [frequency value missing]. The server packages the calculation results and the corresponding equipment control parameters for other materials into a single proportioning execution instruction conforming to the Modbus-TCP protocol. This instruction message contains the PLC's IP address, a write-to-a-register function code, the register address corresponding to the cement inverter's frequency, and the data value 212 (assuming the frequency value is amplified 100 times and transmitted as an integer). The instruction is sent to the field PLC via the network. Upon receiving the instruction, the PLC drives the cement screw feeder's inverter to operate at a frequency of 2.12Hz, thereby accurately realizing the optimized material proportions obtained by the system and completing a full closed-loop control from macro-level decision-making to micro-level execution.

[0105] See appendix Figure 4 The present invention also proposes a filling cost decision system based on neural network multi-objective optimization, which includes the following modules: The data preprocessing module is used to acquire time series data and external economic dynamics instructions corresponding to the filling operation, and to merge and splice the time series data and external economic dynamics instructions to generate basic data sequences. The economic gradient parsing module is used to parse the historical material cost data stream contained in the basic data sequence to calculate the evolutionary evaluation quantity, and generate the economic optimization gradient feature stream based on the evolutionary evaluation quantity. The performance drift analysis module is used to extract the compressive strength test values ​​fed back from the basic data sequence for offset comparison and to calculate and generate fatigue shrinkage penalty variables; The constraint calculation module is used to load the physical state parameters contained in the basic data sequence into the preset physical model, introduce fatigue shrinkage penalty variables to perform boundary shrinkage calculation, and generate safe boundary thresholds. The fusion and update module is used to derive the candidate allocation matrix through the control node based on the pre-trained neural network, combine the economic optimization gradient feature flow with the original detection correction amount output by the neural network to determine the joint optimization gradient, and perform boundary screening on the joint optimization gradient according to the safety boundary threshold to generate controlled update results. The instruction execution module is used to extract the steady-state variable set and convert it into material proportion quota when the controlled update result reaches a stable state. Based on the material proportion quota, it generates proportion execution instructions and sends them to the controlled equipment.

[0106] 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.

[0107] 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 filling cost decision-making method based on neural network multi-objective optimization, characterized in that, Includes the following steps: S1, acquire the time series data and external economic dynamics instructions corresponding to the filling operation, and merge and splice the time series data and external economic dynamics instructions to generate a basic data sequence; S2, parse the historical material cost data stream contained in the basic data sequence to calculate the evolutionary evaluation quantity, and generate the economic optimization gradient feature stream based on the evolutionary evaluation quantity; S3, extract the compressive strength test values ​​from the basic data sequence feedback for offset comparison, and calculate and generate fatigue shrinkage penalty variables; S4, Load the physical state parameters contained in the basic data sequence into the preset physical model, introduce fatigue shrinkage penalty variables to perform boundary shrinkage calculation, and generate a safe boundary threshold. S5. The candidate allocation matrix is ​​derived through the control node based on the pre-trained neural network. The joint optimization gradient is determined by combining the economic optimization gradient feature flow and the original detection correction amount output by the neural network. The joint optimization gradient is then filtered according to the safety boundary threshold to generate a controlled update result. S6. When the controlled update result reaches a stable state, extract the steady-state variable group and convert it into a material proportion quota. Generate a proportion execution instruction based on the material proportion quota and send it to the controlled equipment.

2. The filling cost decision method based on neural network multi-objective optimization according to claim 1, characterized in that, Step S1 includes the following steps: Periodically read tailings concentration characteristic data, pipeline filling differential pressure data, and electromagnetic flow data to form time series data; By invoking material price change signals and batch planned consumption, external economic dynamics instructions are formed. External economic dynamics commands are upsampled according to a preset uniform sampling frequency to align with the time axis of the time series data; The upsampled external economic dynamics instructions are concatenated with time series data to generate a basic data sequence.

3. The filling cost decision method based on neural network multi-objective optimization according to claim 1, characterized in that, Step S2 includes the following steps: Extract material price sampling points within a time window from the basic data sequence to construct a historical data stream of material costs; The price change acceleration is obtained by performing second-order difference calculation on the historical data stream of material costs. The price change acceleration is converted into a penalty coefficient matrix using a nonlinear mapping function; Multiplying the penalty coefficient matrix by the baseline gradient vector yields the economic optimization gradient feature flow.

4. The filling cost decision method based on neural network multi-objective optimization according to claim 1, characterized in that, Step S3 includes the following steps: Receive compressive strength test values ​​with timestamps; By using timestamps, the flow rate decay value corresponding to the production period can be traced back in the basic data sequence; The comprehensive drift measure level is calculated based on the strength deviation of the compressive strength test value relative to the design threshold and the flow velocity decay value. The fatigue shrinkage penalty variable is calculated and generated based on the comprehensive drift level and the preset tolerance.

5. The filling cost decision method based on neural network multi-objective optimization according to claim 1, characterized in that, Step S4 includes the following steps: Input the current slurry ratio information and friction loss parameters from the basic data sequence into the physical model; The initial boundary is a pre-defined fluid dynamics boundary. The initial boundary is adjusted to obtain the dynamic shrinkage boundary; Based on the dynamic contraction boundary, the critical yield stress and ultimate plastic viscosity are output in the physical model to form a safety boundary threshold.

6. The filling cost decision method based on neural network multi-objective optimization according to claim 1, characterized in that, Step S5 includes the following steps: Generate a candidate ratio matrix based on the current ratio parameters; The joint optimization gradient is obtained by weighted summation of the original detection correction amount and the economic optimization gradient feature flow. Predict the physical parameters of each candidate scheme in the candidate ratio matrix; Determine whether each predicted physical parameter exceeds the safety boundary threshold; A usable mask is generated for the candidate solutions that have not exceeded the limit, and the mask is multiplied with the joint optimization gradient to derive the controlled update processing result.

7. The filling cost decision method based on neural network multi-objective optimization according to claim 1, characterized in that, Step S6 includes the following steps: Calculate the parameter update amount of the controlled update result in successive iterations; When the parameter update amount is lower than the preset convergence threshold, it is confirmed that a stable state has been reached, and the steady-state variable set is extracted. Extract the required material proportion quotas for each material from the steady-state variable set; According to the equipment interaction protocol, the material proportioning quota is converted into proportioning execution instructions and sent to the controlled equipment through the communication bus.

8. The filling cost decision method based on neural network multi-objective optimization according to claim 2, characterized in that, Before aligning the time axis of the time series data, the following steps are also included: Identify and read high-frequency noise in pipeline filling differential pressure data; Wavelet transform is applied to filter and remove wavelet coefficients corresponding to high-frequency noise; The wavelet coefficients are reconstructed to obtain the denoised filling differential pressure data.

9. The filling cost decision method based on neural network multi-objective optimization according to claim 4, characterized in that, Calculating the overall drift measure includes: Calculate the ratio of the difference between the test compressive strength value and the design target strength; Calculate the portion of the absolute value of the flow velocity decay value that exceeds the absolute value of the reference flow velocity decay rate, and determine the percentage difference accordingly. The weighted sum of the two difference ratios yields the overall drift measure.

10. The filling cost decision method based on neural network multi-objective optimization according to claim 6, characterized in that, Determine whether each predicted physical parameter exceeds the safety boundary threshold, including: The predicted yield stress and plastic viscosity of each candidate scheme are compared with the corresponding upper and lower limits in the safety boundary threshold. If any predicted value exceeds the corresponding upper or lower limit range, the mask of the candidate solution is set to zero.

Images