A method and system for analyzing and optimizing the cost level of a typical factory prefabricated foundation

Abstract

This invention discloses a method and system for analyzing and optimizing the cost level of typical prefabricated foundations, relating to the fields of engineering cost and intelligent construction technology. The method includes the following steps: setting multiple monitoring objectives based on factors influencing cost levels; optimizing a pre-set first model using a training algorithm based on loss surface sharpness constraints to obtain a second model; outputting a set of monitoring objective scores through the second model; performing scalar evaluation on the monitoring objective score set to generate a compromise solution set; filtering the compromise solution set according to dominance relationship judgment rules and constraining it based on the monitoring objectives to generate an optimization scheme. This application solves the problems of insufficient model robustness and the inability to balance and optimize multiple objectives in cost analysis in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of engineering cost and intelligent construction technology, and more specifically, to a method and system for analyzing and optimizing the cost level of typical prefabricated foundations. Background Technology

[0002] As the application of prefabrication in basic engineering projects such as power, municipal and communications continues to expand, and the factors involved, such as materials, transportation, hoisting and construction teams, become increasingly diversified, projects need to comprehensively process multi-source time-series data during cost assessment and scheme selection. In order to achieve cost control and risk control, the industry is gradually adopting data-driven forecasting and decision-making methods to quantitatively assess different construction scenarios and try to use the assessment results to guide procurement, production scheduling and site entry.

[0003] However, existing cost analysis and scenario optimization methods still have the following shortcomings: On the one hand, under conditions of multi-source heterogeneity, inconsistent timeliness, and a large number of noisy samples, conventional empirical risk minimization training is sensitive to anomalies / pseudo-labels, and is prone to unstable convergence direction, poor robustness to data drift, and insufficient generalization, which in turn affects the stability and repeatability of subsequent optimization processes; on the other hand, existing solutions mostly adopt single objectives (such as minimum cost) or linear weighting with fixed weights, which makes it difficult to characterize the trade-off relationship between cost, schedule, and risk, and lacks systematic screening and solution set generation based on dominance relationships, resulting in results that are highly sensitive to weight settings and making it difficult to obtain a compromise solution set with comparability and consistency. Summary of the Invention

[0004] To overcome the aforementioned deficiencies in the prior art, embodiments of the present invention provide a typical method and system for analyzing and optimizing the cost level of prefabricated factory foundations. By introducing multiple supervised objectives, a training algorithm based on loss surface sharpness constraints, dominance relationship determination rules, and a multi-task temporal deep learning model, the present invention addresses the problems of insufficient model robustness during the optimization process and the inability of cost analysis to balance and optimize multiple objectives in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A typical method for analyzing and optimizing the cost level of prefabricated foundations includes the following steps: setting multiple monitoring objectives based on factors influencing cost level; optimizing a pre-set first model using a training algorithm based on loss surface sharpness constraints to obtain a second model; outputting a set of monitoring objective scores through the second model; performing scalar evaluation on the set of monitoring objective scores to generate a compromise solution set; filtering the compromise solution set according to the dominance relationship determination rule and constraining it based on the monitoring objectives to generate an optimization scheme.

[0007] In a preferred embodiment, the step of optimizing the preset first model using a training algorithm based on loss surface sharpness constraints to obtain the second model specifically involves: within each training batch, calculating sample reputation scores based on a pre-constructed set of sample segments; calculating sample-level gradients for the first model parameters based on the sample reputation scores, and weighting and aggregating the sample-level gradients to obtain the parameter perturbation direction; calculating the scaling factor of the perturbation radius based on the sample reputation scores and scaling the perturbation radius; generating temporary parameters according to the parameter perturbation direction and the scaled perturbation radius; recalculating the sample-level gradients for the first model parameters under the temporary parameters; and iteratively updating the first model parameters to obtain the second model.

[0008] In a preferred embodiment, the step of calculating the sample reputation score based on the pre-constructed sample segment set specifically involves: calculating the integrity index, update frequency matching index, and latency freshness index for each sample segment in the pre-constructed sample segment set; for each sample segment, normalizing the integrity index, update frequency matching index, and latency freshness index, then combining them according to preset weights and pruning them into intervals to obtain the sample reputation score.

[0009] In a preferred embodiment, the step of calculating the sample-level gradient of the first model parameters based on the sample reputation score and weighting and aggregating the sample-level gradients to obtain the parameter perturbation direction specifically involves: backpropagating the first model parameters based on the composite loss function to obtain the sample-level gradient of each sample segment; using the sample reputation score corresponding to the sample segment as a weighting coefficient to perform a weighted summation of all sample-level gradients to obtain the aggregated gradient; and normalizing the aggregated gradient to obtain the parameter perturbation direction.

[0010] In a preferred embodiment, the step of calculating the scaling factor of the perturbation radius based on the sample reputation score and scaling the perturbation radius specifically involves: calculating the mean sample reputation score of all sample segments; inputting the mean sample reputation score into a preset scaling mapping function to obtain the scaling factor, and performing interval clipping on the scaling factor; and merging the clipped scaling factor with the perturbation radius before scaling to obtain the scaled perturbation radius.

[0011] In a preferred embodiment, the step of outputting the supervision target score set through the second model specifically involves: performing feature construction on the sample segment set to generate a basic input tensor; generating a candidate scenario set based on the discrete value set of preset decision variables; mapping each candidate scenario to a candidate scenario vector and concatenating it with the basic input tensor to obtain an inference input tensor; performing forward inference on each inference input tensor through the second model to obtain the corresponding supervision target score; performing normalization processing on the candidate scenarios to obtain candidate scenario identifiers, and using the candidate scenario identifiers as indexes to aggregate the corresponding supervision target scores to form a supervision target score set.

[0012] In a preferred embodiment, the step of performing scalar evaluation on the set of supervised target scores to generate a compromise solution set specifically involves: constructing a set of multi-target weight vectors, where each multi-target weight vector corresponds to a set of multiple supervised targets; performing scale alignment on the supervised target scores in the set of supervised target scores; under each multi-target weight vector, weighted summing of the scale-aligned set of supervised target scores to obtain a scalar evaluation value; traversing all candidate scenarios under each multi-target weight vector, and establishing a one-to-one correspondence between the candidate scenario identifier corresponding to the candidate scenario with the smallest scalar evaluation value and the current multi-target weight vector; and deduplicating and aggregating the candidate scenario identifiers appearing in the one-to-one correspondence to generate a compromise solution set.

[0013] In a preferred embodiment, the compromise solution set is screened according to the dominance relationship determination rule and constrained according to the supervision target to generate an optimization scheme. Specifically, the compromise solution set is determined according to the dominance relationship determination rule; the dominated candidate scenario identifiers in the compromise solution set are removed according to the determination result, and the dominant candidate scenario identifiers are retained. Combined with the corresponding supervision target score, a non-dominated solution set is formed; the non-dominated solution set is constrained to generate an optimization scheme.

[0014] In a preferred embodiment, the dominance relationship is specifically defined as follows: if the first candidate scenario is not greater than the second candidate scenario in all components and is less than the second candidate scenario in at least one component, then the first candidate scenario dominates the second candidate scenario, the candidate scenario identifier corresponding to the first candidate scenario is used as the dominating candidate scenario identifier, and the candidate scenario identifier corresponding to the second candidate scenario is used as the dominated candidate scenario identifier.

[0015] A typical prefabricated foundation cost level analysis and optimization system includes: a model training module, used to optimize a preset first model through a training algorithm based on loss surface sharpness constraints to obtain a second model; a scoring generation module, used to set multiple monitoring objectives based on cost level influencing factors and output a set of monitoring objective scores through the second model; a scalar evaluation module, used to perform scalar evaluation on the set of monitoring objective scores to generate a compromise solution set; and an optimization screening module, used to screen the compromise solution set according to the dominance relationship determination rule and constrain it according to the monitoring objectives to generate an optimization scheme.

[0016] This invention introduces a training algorithm based on loss surface sharpness constraints, which can reduce the interference of noise and heterogeneous data on the training direction, improve the robustness and generalization ability of the model, and thus improve the stability, comparability and decision quality of the optimization process. It effectively solves the problems of insufficient model robustness and the inability of cost analysis to balance and optimize multiple objectives in the prior art. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a typical method for analyzing and optimizing the cost level of prefabricated foundations in a factory, provided as an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of a typical prefabricated foundation cost level analysis and optimization system provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1, Figure 1 A typical method for analyzing and optimizing the cost level of prefabricated foundations in factories is presented, including the following steps:

[0021] S1, set multiple supervision objectives based on the factors affecting the cost level, and optimize the preset first model through a training algorithm based on the sharpness constraint of the loss surface to obtain the second model;

[0022] S2, outputs a set of supervised target scores through the second model;

[0023] S3, performs scalar evaluation on the set of monitoring target scores to generate a compromise solution set;

[0024] S4: The compromise solution set is filtered according to the dominance relationship determination rule and constrained according to the supervision objective to generate an optimal solution.

[0025] This invention introduces a training algorithm based on loss surface sharpness constraints and combines sample reputation scores to weight sample-level gradients with adaptive scaling of the perturbation radius. This reduces the interference of noise and heterogeneous data on the training direction, improving the robustness and generalization ability of the model. By deterministically encoding candidate construction scenarios and aligning them with the basic input tensor, and relying on the output of a multi-task temporal deep learning model with score vectors corresponding to multiple supervised objectives, a scenario-conditional joint assessment of cost, schedule, and quality / safety risks is achieved. By constructing a set of multi-objective weight vectors to scalarize and evaluate the score set, and combining the dominance relationship judgment to generate a compromise solution, the solution is consistently verified under engineering constraints such as budget upper limits, schedule upper limits, and quality / safety risk thresholds, forming an executable optimization scheme. Thus, the stability, comparability, and decision quality of the optimization process are improved, effectively solving the problems of insufficient model robustness and the inability of cost analysis to balance and optimize multiple objectives in existing technologies.

[0026] S1. Based on the factors affecting the cost level, multiple supervision objectives are set, and the first preset model is optimized by a training algorithm based on the sharpness constraint of the loss surface to obtain the second model.

[0027] It should be noted that the multiple supervision objectives can be configured as follows:

[0028] Cost target: The cost is calculated based on a comprehensive unit price of "production and installation costs + transportation tiered fees"; existing suggested values ​​can be directly adopted or recalculated using the same caliber (e.g., a ring main unit with two inlets and two outlets for 60km is approximately 33,448 yuan / unit; a standard box-type transformer for 60km is approximately 38,870 yuan / unit; a cable trench of 2.3m×2.0m for 60km is approximately 6,686 yuan / m).

[0029] Schedule targets: Measured in total construction days, or in the negative vector of "shortened days" (e.g., the average shortening of ring network box foundation by about 7 days and box-type transformer by about 13 days after prefabrication, which can be used as an upper limit / prior reference).

[0030] Quality / Safety Risk Target: Risk score formed by weighting indicators such as the proportion of high-risk operation time, frequency and tonnage of lifting and hoisting operations, proportion of time exposed to the foundation pit, and peak number of on-site workers / total man-hours.

[0031] In this embodiment, the process of optimizing the preset first model using a training algorithm based on loss surface sharpness constraints to obtain the second model is as follows:

[0032] Within each training batch, sample reputation scores are calculated based on a pre-built set of sample segments;

[0033] The sample-level gradient is calculated based on the sample reputation score for the first model parameters, and the sample-level gradient is weighted and aggregated to obtain the parameter perturbation direction.

[0034] Calculate the scaling factor of the disturbance radius based on the sample reputation score and then scale the disturbance radius accordingly;

[0035] Temporary parameters are generated according to the direction of parameter perturbation and the scaled perturbation radius. The sample-level gradients of the first model parameters are recalculated under the temporary parameters, and the first model parameters are iteratively updated to obtain the second model.

[0036] It should be noted that the first model includes a multi-task temporal deep learning model, wherein the multi-task branch layer in the first model is constructed corresponding to multiple supervision objectives, specifically as follows:

[0037] The cost branch corresponds to the cost target;

[0038] The progress branches correspond to the progress targets;

[0039] The quality / safety risk branch corresponds to the quality / safety risk objective.

[0040] It should be noted that the generation of temporary parameters according to the parameter perturbation direction and the scaled perturbation radius specifically refers to:

[0041] Without overwriting the original parameters, create a tensor for each trainable parameter. Generate a temporary parametric tensor of the same shape. And perform addition element by element:

[0042]

[0043] in, This is the scaled perturbation radius. for The unit perturbation direction component on, and Consistent shape; all Combine them to obtain temporary parameters.

[0044] It should be noted that the first model parameter refers to the set of trainable parameters that participate in backpropagation and are updated by the optimizer during the training process.

[0045] In this embodiment, the calculation of sample reputation score based on the pre-constructed sample segment set specifically involves:

[0046] For each sample segment in the pre-constructed sample segment set, calculate the integrity index, update frequency matching index, and latency freshness index.

[0047] For each sample segment, the integrity index, update frequency matching index, and time delay freshness index are normalized, weighted and combined according to preset weights, and then pruned by interval to obtain the sample reputation score.

[0048] It should be noted that the pre-built sample segment set is constructed with "project-module-time period" as the granularity, and then the granularity is aligned. Each sample segment contains working condition data, monitoring target availability label and time data. The working condition data includes price, shift, transportation and site conditions, and the time data includes timestamp, update record, collection and storage time.

[0049] It should be noted that the integrity index is specifically as follows:

[0050] The integrity index is used to evaluate the completeness of fields and time series. Preferably, the integrity index can be calculated in the following form:

[0051] Calculate the integrity of the working condition data: Perform availability determination on each category of the working condition data one by one. The determination rule is based on the time window of the sample segment. If the effective record coverage of a certain category of data is not lower than the preset threshold, then the data of that category is determined to be available. The ratio of the number of data in each category that is determined to be available to the number of required data is used as the integrity of the working condition data of the sample segment.

[0052] Recalculate timing integrity:

[0053]

[0054] in, To ensure the temporal integrity of the sample segment, The desired number of time steps;

[0055] Recalculate tag availability:

[0056]

[0057] in, For the label availability of the sample segment, This is an indicator function; it takes the value 1 if all the supervision target labels are available, and 0 otherwise.

[0058] Finally, the data integrity, temporal integrity, and tag availability were weighted and summed after softmax normalization:

[0059]

[0060] in, This is an indicator of the integrity of the sample segment. To ensure the integrity of the operating condition data in the sample segment, To ensure the temporal integrity of the sample segment, For the label availability of the sample segment, These are weighted coefficients for field integrity, temporal integrity, and tag availability, respectively, and the weighted coefficients can be set with fixed weights by business priors.

[0061] It should be noted that the update frequency matching index is specifically as follows:

[0062] The update frequency matching index is used to measure the actual update rhythm and the expected update rhythm of dynamic data fields within a sample segment. Preferably, the completeness index can be calculated in the following form:

[0063] The update frequency matching index is calculated based on the relative deviation.

[0064]

[0065] in, The update frequency matching index for the sample segment. This represents the actual update cycle of the dynamic data fields within the sample segment. This represents the expected update cycle for dynamic data fields within the sample segment. This is the tolerance coefficient, and its value range is [0.1, 1]. The larger the size, the higher the tolerance.

[0066] It should be noted that the aforementioned time-delay freshness index specifically refers to:

[0067] The time-delay freshness index is used to measure the interval between the last update time of a dynamic data field within a sample segment and the current time. Preferably, the time-delay freshness index can be calculated in the following form:

[0068] Calculate the time-latency freshness index based on exponential decay:

[0069]

[0070] in, The time delay freshness index for the sample segment This represents the interval between the last update time of a dynamic data field and the current time. It is a time constant and is greater than 0. The larger the time constant, the slower the decay.

[0071] It should be noted that the process of weighted summation according to preset weights and interval pruning to obtain the sample reputation score is as follows:

[0072] The integrity index, update frequency matching index, and latency freshness index of the current sample segment are weighted and summed according to preset weights, and then cropped to the interval [0,1].

[0073]

[0074] in, For sample segments Sample reputation score, Indicates the result Truncate to [0,1] These are preset weights for the integrity index, update frequency matching index, and latency freshness index, respectively, and the preset weights are non-negative and Fixed weights can be set by business prior knowledge. This is the integrity index after normalization of the sample segment. The update frequency matching index is the normalized update frequency of the sample segment. This is the time delay freshness index after normalization of the sample segment.

[0075] In this embodiment, the step of calculating the sample-level gradient of the first model parameters based on the sample reputation score, and then weighting and aggregating the sample-level gradients to obtain the parameter perturbation direction, specifically involves:

[0076] Backpropagation of the first model parameters is performed based on the composite loss function to obtain the sample-level gradient for each sample segment;

[0077] The sample reputation score corresponding to the sample segment is used as the weighting coefficient, and the gradients of all sample levels are weighted and summed to obtain the aggregate gradient.

[0078] Normalize the aggregated gradient to obtain the direction of parameter perturbation.

[0079] It should be noted that the composite loss function is composed of the single-target losses of multiple supervised targets:

[0080] First, the single objective of multiple supervision objectives is uniformly minimized by monotonic transformation and scale alignment, and normalized to [0,1]. The cost objective adopts logarithmically compressed Huber loss, the schedule objective adopts Huber loss, and the quality / safety risk objective selects weighted binary cross-entropy or Brier loss according to the label type.

[0081] Preferably, the composite loss function is constructed using a fixed weighted sum:

[0082]

[0083] in, For compound loss, The single-objective loss corresponding to the cost target. The single-objective loss corresponding to the schedule target. For the single-objective loss corresponding to the quality / safety risk objective, The weights corresponding to the single-objective loss of the three supervision objectives mentioned above can be tuned through business preferences or grid / validation sets.

[0084] It should be noted that the backpropagation of the first model parameters based on the composite loss function to obtain the sample-level gradient for each sample segment is specifically as follows:

[0085] Within the current training batch, calculate the composite loss for each sample segment based on the composite loss function;

[0086] For each sample segment, perform a backpropagation on the parameters of the first model with the composite loss as the target, obtain the sample-level gradient of the sample segment and cache it, and perform this process sequentially on each sample segment in the current training batch.

[0087] In this embodiment, the step of calculating the scaling factor of the perturbation radius based on the sample reputation score and scaling the perturbation radius specifically involves:

[0088] Calculate the mean sample reputation score for all sample segments;

[0089] Input the mean of the sample reputation score into a preset scaling mapping function to obtain the scaling coefficient, and then perform interval clipping on the scaling coefficient;

[0090] The scaled perturbation radius is obtained by combining the clipped scaling factor with the original perturbation radius.

[0091] It should be noted that the preset scaling mapping function is specifically as follows:

[0092]

[0093] in, This is the scaling factor. The mean of the sample reputation. , These are breakpoints used only to determine the start / end of a linear transition, and ; This is the lower bound of the scaling factor. This is the upper bound of the scaling factor.

[0094] It should be noted that the aforementioned range clipping of the scaling factor refers to applying a constraint within a range to the scaling factor to avoid numerical anomalies, ensuring that the scaling factor always remains within a certain range. Inside.

[0095] It should be noted that the process of merging the cropped scaling factor with the original perturbation radius specifically involves multiplying the cropped scaling factor by the original perturbation radius.

[0096] S2 outputs a set of supervised target scores through the second model.

[0097] In this embodiment, the step of outputting the supervision target score set through the second model specifically refers to:

[0098] Perform feature construction on the sample set to generate the basic input tensor;

[0099] And generate a set of candidate scenarios based on the discrete value set of the preset decision variables;

[0100] Each candidate scenario is mapped to a candidate scenario vector, and then concatenated with the basic input tensor to obtain the inference input tensor.

[0101] The second model performs forward inference on each inference input tensor to obtain the corresponding supervised target score.

[0102] Candidate scenarios are normalized to obtain candidate scenario identifiers, and the corresponding supervision target scores are collected using the candidate scenario identifiers as indexes to form a supervision target score set.

[0103] It should be noted that the process of constructing features on the sample segment set to generate the basic input tensor specifically involves: extracting original fields from the pre-constructed sample segment set at the granularity of "project-module-time period" and aligning the granularity; performing outlier pruning and missing value imputation on the working condition data and time data, and generating derived features (such as unit price × usage, capacity utilization rate, transportation intensity, etc.) according to business rules; performing numerical encoding (ordered encoding or one-hot encoding) on ​​category / discrete fields, performing scale alignment on continuous fields according to preset statistics (such as Z-Score or quantile scaling), and unifying the target direction and dimensions; and stacking them on the feature axis according to the preset feature column order to form the basic input tensor of each sample segment.

[0104] It should be noted that the candidate scenario set is constructed in the following way:

[0105] A predefined set of decision variables is provided, where the discrete value set of each decision variable can be:

[0106] (1) Module limitation: {ring network box / substation / cable well / cable trench / pipeline};

[0107] (2) Transportation distance tiers: {30km, 60km, 90km};

[0108] (3) Loading method: {full vehicle, partial load} etc.;

[0109] Calculate the Cartesian product for the discrete value set of each decision variable to obtain the initial set of candidate scenarios;

[0110] Define a decision function, such as equipment capacity ≥ component weight, road width and weight restrictions, or prohibition of nighttime hoisting, and perform feasibility filtering on the initial set of candidate scenarios to obtain the candidate scenario set.

[0111] It should be noted that the process of performing normalization on candidate scenarios to obtain candidate scenario identifiers specifically involves unifying the field order, unit of measurement and decimal places, discrete values ​​and quantization step size, and outlier pruning according to a preset configuration table to form normalized tuples, which are then serialized into normalized strings using a fixed delimiter. The normalized strings are directly used as candidate scenario identifiers.

[0112] S3 performs scalar evaluation on the set of scores for the supervised targets, generating a compromise solution set.

[0113] In this embodiment, the step of performing scalar evaluation on the set of supervised target scores to generate a compromise solution set specifically involves:

[0114] Construct a multi-objective weight vector set, wherein each multi-objective weight vector in the multi-objective weight vector set corresponds to a set of multiple supervision objectives;

[0115] Align the performance scales of the supervisory target scores within the supervisory target score set;

[0116] Under each multi-objective weight vector, the weighted summation of a set of supervised objective scores after scale alignment is obtained to obtain the scalarized evaluation value.

[0117] Under each multi-objective weight vector, all candidate scenarios are traversed, and for the candidate scenario with the smallest quantification evaluation value, a one-to-one correspondence is established between the corresponding candidate scenario identifier and the current multi-objective weight vector.

[0118] The candidate scenario identifiers appearing in the one-to-one correspondence are deduplicated and aggregated to generate a compromise solution set.

[0119] It should be noted that the construction of the multi-objective weight vector set can be achieved using an equally spaced simplex mesh, specifically as follows:

[0120] Determine the resolution as ,and And stipulate the correspondence between multiple supervision objectives and weight components ( (These correspond to cost targets, schedule targets, and quality / safety risk targets, respectively).

[0121] Definition satisfies and Discrete integer grid, taking non-negative integers. make Corresponding weight ;

[0122] The weights are standardized in terms of numerical values, with the weights expressed in rational number steps. express;

[0123] Output multi-objective weight vector set .

[0124] S4: The compromise solution set is filtered according to the dominance relationship determination rule and constrained according to the supervision objective to generate an optimal solution.

[0125] In this embodiment, the compromise solution set is screened according to the dominance relationship determination rule and constrained according to the supervision objective to generate an optimization scheme, specifically as follows:

[0126] The compromise solution set will be determined according to the dominance relationship determination rule;

[0127] Based on the judgment results, the dominated candidate scenario identifiers in the compromise solution set are removed, and the dominated candidate scenario identifiers are retained. Combined with the corresponding supervision target scores, a non-dominated solution set is formed.

[0128] Constraints are applied to the non-dominated solution set to generate an optimal solution.

[0129] It should be noted that the constraint processing specifically refers to:

[0130] Set a cost target threshold, optionally taking the contract / approved budget or control price; set a schedule target threshold, optionally taking the milestone / commissioning period; set a quality / safety risk target threshold, optionally taking the enterprise's EHS red line score; based on the above thresholds, constrain the monitoring target score corresponding to each candidate scenario in the non-dominated solution set, and eliminate candidate scenarios and their monitoring target scores that do not meet the constraints in the non-dominated solution set;

[0131] Furthermore, the set of constraints includes at least thresholds corresponding to multiple monitoring targets, and may also include other engineering feasibility constraints, such as equipment capacity, road width and weight limits, work group limits, and nighttime lifting bans.

[0132] It should be noted that the optimization scheme includes candidate scenario identification and supervision target scoring;

[0133] Furthermore, the optimization scheme is optimized into an action list, which can identify the candidate scenarios corresponding to the candidate scenarios and restore them to executable parameters using the inverse mapping of a preset deterministic encoding function.

[0134] In this embodiment, the dominance relationship is specifically defined as follows: if the first candidate scenario is not greater than the second candidate scenario in all components and is less than the second candidate scenario in at least one component, then the first candidate scenario dominates the second candidate scenario, the candidate scenario identifier corresponding to the first candidate scenario is used as the dominating candidate scenario identifier, and the candidate scenario identifier corresponding to the second candidate scenario is used as the dominated candidate scenario identifier.

[0135] This application combines sample reputation scores with weighted gradients at the sample level and adaptive scaling of the perturbation radius to reduce the interference of noise and heterogeneous data on the training direction, thereby improving the robustness and generalization ability of the model. By deterministically encoding candidate construction scenarios and aligning them with the basic input tensor, and relying on the output of a multi-task temporal deep learning model to achieve a scenario-conditional joint assessment of cost, schedule, and quality / safety risks, the application constructs a set of multi-objective weight vectors to perform scalar evaluation of the score set and generates a compromise solution by combining the dominance relationship determination. This solution is consistently verified under engineering constraints such as budget limits, schedule limits, and quality / safety risk thresholds, resulting in an executable optimization scheme.

[0136] Example 2, Figure 2 This invention presents a typical prefabricated foundation cost level analysis and optimization system, comprising:

[0137] The model training module is used to optimize the preset first model through a training algorithm based on the loss surface sharpness constraint to obtain the second model;

[0138] The scoring generation module is used to set multiple supervision targets based on factors affecting the cost level, and outputs a set of supervision target scores through the second model.

[0139] The scalarization evaluation module is used to perform scalarization evaluation on the set of scores for supervision targets and generate a set of compromise solutions.

[0140] The optimization and filtering module is used to filter the compromise solution set according to the dominance relationship determination rule, and to generate an optimization scheme based on the supervision objective.

[0141] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0142] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0143] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0144] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0145] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0146] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for analyzing and optimizing the cost level of typical prefabricated foundations in factories, characterized in that, Includes the following steps: Multiple monitoring targets are set according to the factors affecting the cost level. A first model is optimized by a training algorithm based on the sharpness constraint of the loss surface to obtain a second model. The training algorithm includes dynamically weighting the gradient of the first model based on the sample reputation score to determine the perturbation direction and scaling the perturbation radius, thereby generating temporary parameters to iteratively optimize the first model. The second model outputs a set of supervisory target scores. Perform scalar evaluation on the set of monitoring target scores to generate a set of compromise solutions; The compromise solution set is selected according to the dominance relationship determination rule and constrained according to the supervision objective to generate an optimal solution.

2. The method for analyzing and optimizing the cost level of typical prefabricated foundations according to claim 1, characterized in that, The second model is obtained by optimizing the preset first model through a training algorithm based on the sharpness constraint of the loss surface, specifically as follows: Within each training batch, sample reputation scores are calculated based on a pre-built set of sample segments; The sample-level gradient is calculated based on the sample reputation score for the first model parameters, and the sample-level gradient is weighted and aggregated to obtain the parameter perturbation direction. Calculate the scaling factor of the disturbance radius based on the sample reputation score and then scale the disturbance radius accordingly; Temporary parameters are generated according to the direction of parameter perturbation and the scaled perturbation radius. The sample-level gradients of the first model parameters are recalculated under the temporary parameters, and the first model parameters are iteratively updated to obtain the second model.

3. The method for analyzing and optimizing the cost level of typical prefabricated foundations according to claim 2, characterized in that, The calculation of sample reputation scores based on a pre-constructed set of sample segments is specifically as follows: For each sample segment in the pre-constructed sample segment set, calculate the integrity index, update frequency matching index, and latency freshness index. For each sample segment, the integrity index, update frequency matching index, and time delay freshness index are normalized, weighted and combined according to preset weights, and then pruned by interval to obtain the sample reputation score.

4. The method for analyzing and optimizing the cost level of typical prefabricated foundations according to claim 3, characterized in that, The step of calculating the sample-level gradient of the first model parameters based on the sample reputation score, and then weighting and aggregating the sample-level gradients to obtain the parameter perturbation direction, specifically involves: Backpropagation of the first model parameters is performed based on the composite loss function to obtain the sample-level gradient for each sample segment; The sample reputation score corresponding to the sample segment is used as the weighting coefficient, and the gradients of all sample levels are weighted and summed to obtain the aggregate gradient. Normalize the aggregated gradient to obtain the direction of parameter perturbation.

5. The method for analyzing and optimizing the cost level of typical prefabricated foundations according to claim 4, characterized in that, The process of calculating the scaling factor of the perturbation radius based on the sample reputation score and scaling the perturbation radius specifically involves: Calculate the mean sample reputation score for all sample segments; Input the mean of the sample reputation score into a preset scaling mapping function to obtain the scaling coefficient, and then perform interval clipping on the scaling coefficient; The scaled perturbation radius is obtained by combining the clipped scaling factor with the original perturbation radius.

6. The method for analyzing and optimizing the cost level of typical prefabricated foundations according to claim 5, characterized in that, The output of the supervised target score set through the second model is specifically as follows: Perform feature construction on the sample set to generate the basic input tensor; And generate a set of candidate scenarios based on the discrete value set of the preset decision variables; Each candidate scenario is mapped to a candidate scenario vector, and then concatenated with the basic input tensor to obtain the inference input tensor. The second model performs forward inference on each inference input tensor to obtain the corresponding supervised target score. Candidate scenarios are normalized to obtain candidate scenario identifiers, and the corresponding supervision target scores are collected using the candidate scenario identifiers as indexes to form a supervision target score set.

7. The method for analyzing and optimizing the cost level of typical prefabricated foundations according to claim 6, characterized in that, The process of performing scalar evaluation on the set of supervised target scores to generate a compromise solution set is as follows: Construct a multi-objective weight vector set, wherein each multi-objective weight vector in the multi-objective weight vector set corresponds to a set of multiple supervision objectives; Align the performance scales of the supervisory target scores within the supervisory target score set; Under each multi-objective weight vector, the weighted summation of a set of supervised objective scores after scale alignment is obtained to obtain the scalarized evaluation value. Under each multi-objective weight vector, all candidate scenarios are traversed, and for the candidate scenario with the smallest quantification evaluation value, a one-to-one correspondence is established between the corresponding candidate scenario identifier and the current multi-objective weight vector. The candidate scenario identifiers appearing in the one-to-one correspondence are deduplicated and aggregated to generate a compromise solution set.

8. The method for analyzing and optimizing the cost level of typical prefabricated foundations according to claim 7, characterized in that, The compromise solution set is filtered according to the dominance relationship determination rule and constrained according to the supervision objective to generate an optimization scheme, specifically as follows: The compromise solution set will be determined according to the dominance relationship determination rule; Based on the judgment results, the dominated candidate scenario identifiers in the compromise solution set are removed, and the dominated candidate scenario identifiers are retained. Combined with the corresponding supervision target scores, a non-dominated solution set is formed. Constraints are applied to the non-dominated solution set to generate an optimal solution.

9. The method for analyzing and optimizing the cost level of typical prefabricated foundations according to claim 8, characterized in that, The dominance relationship is as follows: if the first candidate scenario is not greater than the second candidate scenario in all components and is less than the second candidate scenario in at least one component, then the first candidate scenario dominates the second candidate scenario, the candidate scenario identifier corresponding to the first candidate scenario is used as the dominating candidate scenario identifier, and the candidate scenario identifier corresponding to the second candidate scenario is used as the dominated candidate scenario identifier.

10. A system using a typical prefabricated foundation cost level analysis and optimization method as described in any one of claims 1-9, comprising: The model training module is used to set multiple supervision objectives based on the factors affecting the cost level, and optimize the preset first model through a training algorithm based on the sharpness constraint of the loss surface to obtain the second model; The scoring generation module is used to output a set of supervised target scores through the second model; The scalarization evaluation module is used to perform scalarization evaluation on the set of scores for supervision targets and generate a set of compromise solutions. The optimization and filtering module is used to filter the compromise solution set according to the dominance relationship determination rule, and to generate an optimization scheme based on the supervision objective.

Images