A research and development cost dynamic optimization method based on multi-source data fusion

Abstract

The application relates to the technical field of data processing, and discloses a research and development cost dynamic optimization method based on multi-source data fusion, which comprises the following steps: acquiring multi-source heterogeneous research and development data to generate a standardized data set; constructing a semantic constraint relationship graph and calculating an uncertainty quantitative value; generating a dynamic fusion weight, weighting and fusing the features of each data source, and recalculating the dynamic fusion weight when the standardized data set or the semantic constraint relationship graph is updated; extracting a key causal path to generate a research and development cost dynamic prediction result; calculating a deviation to perform deviation attribution analysis and generate a deviation feedback signal; based on a semantic level matching result, calling a corresponding optimization strategy or triggering a knowledge discovery process, and updating the edge weight of the semantic constraint relationship graph and the dynamic fusion weight. The application realizes multi-source data adaptive fusion and closed-loop self-optimization, and improves the research and development cost prediction accuracy and the traceability of deviation attribution.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for dynamic optimization of R&D costs based on multi-source data fusion. Background Technology

[0002] R&D projects are characterized by long cycles, high uncertainty, multiple participating entities, and diverse data sources. R&D cost data is scattered across multiple heterogeneous platforms, such as R&D management systems, supply chain systems, financial systems, and technology innovation achievement management systems. These data sources exhibit significant differences in semantic expression, temporal precision, and format specifications.

[0003] In existing technologies, multi-source data fusion strategies generally employ fixed weights or static models. When the data distribution or reliability of the data sources changes, the accuracy of the fusion results is difficult to guarantee. Furthermore, most existing cost prediction methods are geared towards engineering construction or manufacturing scenarios, making it difficult to effectively handle the complex nonlinear coupling relationships between cost drivers in R&D scenarios. They also lack the ability to incorporate data on technological innovation achievements and intellectual property rights into the cost analysis framework. Regarding deviation analysis and optimization, existing technologies typically treat prediction deviations merely as error records, failing to feed the deviation attribution results back to the data fusion and prediction stages to form a closed-loop optimization. The system cannot continuously learn from historical deviations, resulting in low optimization efficiency.

[0004] Therefore, this invention proposes a dynamic optimization method for R&D costs based on multi-source data fusion. Summary of the Invention

[0005] The purpose of this invention is to address the problems in existing technologies, such as the lack of dynamic adjustment capabilities of fusion strategies, unsuitable cost prediction for R&D, and non-closed-loop deviation analysis. To this end, a dynamic optimization method for R&D costs based on multi-source data fusion is proposed.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for dynamic optimization of R&D costs based on multi-source data fusion, comprising the following steps: Step S1: Obtain multi-source heterogeneous R&D data, perform unified identification and preprocessing, and generate a standardized dataset. The multi-source heterogeneous R&D data includes technological innovation achievement data and intellectual property data. Step S2: Semantic hierarchical annotation of features from each data source is performed, time lag mutual information method is used to identify causal relationships between features, semantic constraint relationship graph with causal relationships as directed edges is constructed, and uncertainty quantification value of each feature is calculated. Step S3: Based on the structural centrality and uncertainty quantification values ​​of each feature in the semantic constraint relationship graph, generate dynamic fusion weights, perform weighted fusion of features from each data source, and recalculate the dynamic fusion weights of each feature when the standardized dataset or semantic constraint relationship graph is updated. Step S4: Extract key causal paths pointing to the cost prediction target from the semantic constraint relationship graph, construct a long short-term memory network prediction model with the features on the key causal paths as attention masks, and generate dynamic prediction results of R&D costs. Step S5: Calculate the deviation between the actual cost data and the prediction results. Based on the interpretability analysis results of the Long Short-Term Memory Network prediction model, perform deviation attribution analysis along the causal path and semantic hierarchy in the semantic constraint relationship graph to generate a deviation feedback signal. Step S6: Based on the semantic hierarchical matching results of the current deviation feedback signal and the historical deviation pattern, call the corresponding optimization strategy or trigger the knowledge discovery process, and update the edge weights and dynamic fusion weights of the semantic constraint relationship graph.

[0007] The beneficial effects of the technical solution provided by this invention include at least the following: This invention, by constructing a semantic constraint relationship graph and calculating the uncertainty quantification value of each feature, can eliminate semantic differences and data quality differences between multi-source heterogeneous data, and achieve adaptive fusion of multi-source R&D data.

[0008] This invention generates dynamic fusion weights based on the structural centrality and uncertainty quantification values ​​in the semantic constraint graph. It can automatically adjust the fusion strategy when the standardized dataset or semantic constraint graph is updated, thereby improving the accuracy of the fusion results.

[0009] This invention extracts key causal paths from semantic constraint graphs and constructs a long short-term memory network prediction model constrained by causal path attention mask. This enables the prediction model to focus only on features that are causally related to cost objectives, thereby improving the accuracy of R&D cost prediction.

[0010] This invention performs deviation attribution analysis along the causal path and semantic hierarchy in the semantic constraint relationship graph. It can trace the transmission chain of deviation in the causal network and locate the source of deviation according to the semantic hierarchy, providing accurate attribution basis for optimization decision-making.

[0011] This invention constructs a deviation pattern knowledge base and matches historical deviation patterns based on semantic hierarchy. It can reuse historical optimization strategies and form a closed-loop self-optimization by updating the edge weights and dynamic fusion weights of the semantic constraint relationship graph. This allows the fusion strategy to continuously evolve with actual deviation feedback, thereby achieving dynamic optimization of R&D costs. Attached Figure Description

[0012] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of the method flow provided in an embodiment of the present invention; Figure 2 A flowchart of the prediction model provided in an embodiment of the present invention. Detailed Implementation

[0014] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a dynamic optimization method for R&D costs based on multi-source data fusion proposed according to the present invention. In the following description, different embodiments or different embodiments do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0015] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0016] The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0017] The following description, in conjunction with the accompanying drawings, details a specific scheme for a dynamic optimization method for R&D costs based on multi-source data fusion provided by this invention.

[0018] Please see Figure 1 The diagram illustrates a flowchart of a method for dynamically optimizing R&D costs based on multi-source data fusion, according to an embodiment of the present invention, including the following steps: Step S1: Obtain multi-source heterogeneous R&D data, perform unified identification and preprocessing, and generate a standardized dataset. Multi-source heterogeneous R&D data includes technological innovation achievement data and intellectual property data. Step S2: Semantic hierarchical annotation of features from each data source is performed, time lag mutual information method is used to identify causal relationships between features, semantic constraint relationship graph with causal relationships as directed edges is constructed, and uncertainty quantification value of each feature is calculated. Step S3: Based on the structural centrality and uncertainty quantification values ​​of each feature in the semantic constraint relationship graph, generate dynamic fusion weights, perform weighted fusion of features from each data source, and recalculate the dynamic fusion weights of each feature when the standardized dataset or semantic constraint relationship graph is updated. Step S4: Extract key causal paths pointing to the cost prediction target from the semantic constraint relationship graph, construct a long short-term memory network prediction model with the features on the key causal paths as attention masks, and generate dynamic prediction results of R&D costs. Step S5: Calculate the deviation between the actual cost data and the prediction results. Based on the interpretability analysis results of the Long Short-Term Memory Network prediction model, perform deviation attribution analysis along the causal path and semantic hierarchy in the semantic constraint relationship graph to generate a deviation feedback signal. Step S6: Based on the semantic hierarchical matching results of the current deviation feedback signal and the historical deviation pattern, call the corresponding optimization strategy or trigger the knowledge discovery process, and update the edge weights and dynamic fusion weights of the semantic constraint relationship graph.

[0019] It should be noted that multi-source heterogeneous R&D data refers to bill of materials data, component purchase price data, process route time data, equipment depreciation data, labor cost data, market material price index data, technological innovation achievement data, and intellectual property data obtained from multiple heterogeneous systems. The data sources differ in terms of time accuracy, semantic expression, and format specifications.

[0020] Technological innovation achievement data refers to the registration information of technological achievements, maturity level of achievements, records of achievement reuse, estimated benefits of achievement transformation, and information on R&D projects related to achievements from the technological innovation achievement sharing service system and the technology achievement transformation and promotion management system.

[0021] Intellectual property data refers to the patent legal status, patent citation relationships, trademark registration status, technical standard relevance, and intellectual property value assessment level from the intelligent patent and trademark precise retrieval and identification system and the digital enterprise credit evaluation platform.

[0022] Unified identification refers to mapping fields from various data sources to standard fields according to a unified data dictionary, and adding a timestamp and data source system identifier to each data record.

[0023] Semantic hierarchical annotation refers to dividing features into four semantic levels—structural features, dynamic features, external features, and outcome features—based on data source and semantic category, and defining directed constraint relationships between each semantic level.

[0024] The time-lag mutual information method refers to calculating the mutual information of any two features under different lag steps, taking the lag direction corresponding to the maximum mutual information value as the causal direction, and establishing a directed edge from cause to effect if the mutual information value is greater than a preset causal threshold, using the normalized mutual information value as the initial edge weight.

[0025] A semantic constraint relationship graph is a directed acyclic graph structure with features at each semantic level as nodes and causal relationships between features as directed edges. The edge types in the graph include edges from structural feature nodes to dynamic feature nodes, edges from external feature nodes to dynamic feature nodes, edges from dynamic feature nodes to outcome feature nodes, and edges from outcome feature nodes to structural feature nodes.

[0026] Uncertainty quantification refers to the measure of the credibility of various characteristic data. Structural characteristics are calculated based on the half-width of the statistical interval, dynamic characteristics are calculated based on the standard deviation of historical volatility, external characteristics are calculated based on the root mean square of the prediction error, and outcome characteristics are calculated based on the confidence interval width of the conversion benefit prediction.

[0027] Structural centrality is an index that reflects the topological importance of a feature in a causal network, calculated by extracting the out-degree and in-degree of each feature from the semantic constraint graph. Features with larger out-degrees have a wider range of influence on downstream features, while features with larger in-degrees are more constrained by upstream features.

[0028] Dynamic fusion weights refer to the weight coefficients of each feature in the weighted fusion calculated by integrating structural centrality and uncertainty quantification values. They are recalculated when a new standardized dataset is generated or the edge weights of the semantic constraint graph are updated.

[0029] The critical causal path refers to all causal paths that point to the cost prediction target node extracted from the semantic constraint relationship graph.

[0030] Attention masking constraint refers to the process in the causal path attention layer where the attention weights of features located on key causal paths are retained, while the attention weights of features not on key causal paths are reset to zero, so that the prediction model makes predictions only based on features on key causal paths.

[0031] Deviation attribution analysis refers to the process of using interpretability analysis tools to obtain the contribution of each input feature in the prediction model to the prediction result when the deviation between the actual cost data and the prediction result exceeds a preset deviation threshold. The contribution is then propagated layer by layer to the cost prediction target node along the causal path in the semantic constraint relationship graph. Based on the cumulative contribution of each semantic level during the propagation process, the feature with the highest cumulative contribution is determined as the contributing feature of that semantic level. The cumulative contribution of each semantic level is compared, and the deviation type corresponding to the semantic level with the largest cumulative contribution is taken as the deviation type. When the cumulative contribution of the outcome feature is the largest, it is marked as an outcome-type deviation.

[0032] Deviation feedback signals refer to information that includes deviation values, deviation types, and contribution characteristics at each semantic level.

[0033] The deviation pattern knowledge base is a database used to store historical deviation feedback signals and their corresponding optimization strategies. Each record includes deviation type, deviation magnitude, contribution characteristics of each semantic level, snapshot of semantic constraint relationship graph, optimization strategy adopted, and evaluation index of optimization effect.

[0034] Semantic hierarchical matching refers to weighted matching that comprehensively considers the semantic hierarchy of deviation type, deviation amplitude range, and contribution features when calculating the similarity between the current deviation feedback signal and each historical deviation pattern in the knowledge base.

[0035] The knowledge discovery process refers to recalculating the time-lag mutual information of the contribution features in the current deviation feedback signal and updating the corresponding causal edge weights.

[0036] In one specific implementation, taking the product R&D cost management of an equipment manufacturing enterprise as an application scenario, the enterprise has deployed multiple information systems such as R&D expense collection, project progress tracking, sharing of technological innovation achievements, and patent and trademark retrieval. The specific implementation process of the present invention is described with an 18-month cycle.

[0037] The specific implementation of step S1 is as follows: Using ETL tools, extract material bill of materials data, component procurement price data, process route time data, labor cost data, technology achievement registration information, achievement reuse records, patent legal status, and patent citation relationship data from the aforementioned systems. Simultaneously, obtain market material price indices from a third-party data platform. The initial value for the technology achievement reuse rate is 45%, meaning that 54 out of the 120 technological achievements registered by the company are cited in subsequent R&D projects. The achievement transformation yield is 120%, meaning the return on achievement transformation is 1.2 times the corresponding R&D investment. Patent legal status includes three states: valid authorization, pending examination, and expired. Patent citation relationships include the number of prior patents cited by this patent and the number of times it is cited by subsequent patents.

[0038] Fields from various data sources are mapped to standard fields using a unified data dictionary, and each data record is timestamped and identified by its source system. For example, a component purchase price record is timestamped as March 15, 2025, 09:30 and identified as originating from the ERP system; a technology reuse record is timestamped as March 15, 2025, 10:00 and identified as originating from the Technology Innovation Achievement Sharing Service System. Missing values ​​are filled using a time-weighted interpolation algorithm, and data segments with consecutive missing values ​​exceeding a preset threshold are filled using the historical average for the same period. Outliers are detected using a sliding window algorithm, and the three-standard-deviation criterion is used to identify anomalies. A standardized dataset is generated, containing fields such as bill of materials data, component purchase price data, process route time data, equipment depreciation data, labor cost data, market material price index data, technology reuse rate, technology transfer yield, patent legal status, and patent citation relationships. Data collection is performed incrementally daily, and a new standardized dataset is generated at the beginning of each month.

[0039] Step S2 is implemented as follows: Semantic hierarchical annotation is performed on the features of each data source in the standardized dataset. The bill of materials (BOM) hierarchy and standard process routes are labeled as structural features. Specifically, the BOM hierarchy represents the four-level structure of a product model: assembly, subassemblies, components, and parts. The standard process route represents the five-step sequence of material preparation, rough machining, finish machining, heat treatment, and assembly. The weekly fluctuation sequence of purchase prices and the daily variation sequence of labor consumption are labeled as dynamic features. Specifically, the purchase price fluctuation represents the week-on-week change rate of the purchase unit price of a certain type of steel over the past 12 weeks. The labor consumption variation represents the deviation rate of the daily actual labor hours from the standard labor hours for the finish machining process over the past 30 days. The steel price index and the exchange rate midpoint are labeled as externalities. Specifically, the steel price index represents the monthly price index of cold-rolled steel sheets in the East China market, and the exchange rate midpoint represents the daily midpoint of the USD / CNY exchange rate. The reuse rate of technological achievements, the rate of return on intellectual property conversion, and the correlation between achievements and R&D projects are labeled as outcome-related features. Specifically, the correlation between achievements and R&D projects refers to the matching degree between a technological achievement and the technical objectives of its source R&D project, with a value between 0 and 1. In this embodiment, the initial value for this feature is 0.75. A total of 25 features are included, comprising 12 structural features, 8 dynamic features, 4 externality features, and 5 outcome-related features.

[0040] Define the directed constraints between each semantic level. Structural features constrain the value range of dynamic features; that is, the bill of materials hierarchy constrains the reasonable fluctuation range of component procurement prices, and standard process routes constrain the upper and lower limits of the quota for labor hours consumed in each process. Externalities influence the evolution direction of dynamic features; that is, the steel price index affects the long-term evolution direction of steel procurement prices, and the exchange rate midpoint affects the changing trend of imported component procurement costs. Dynamic features affect the output efficiency of outcome features; that is, fluctuations in procurement prices and changes in labor hours jointly affect the output efficiency of technological achievements. Outcome features, in turn, feed back into adjustments to structural features; that is, the reuse rate of technological achievements, the rate of return on intellectual property conversion, and the relevance of achievements to R&D projects feed back into the optimization of the bill of materials and the improvement of standard process routes.

[0041] For each of the 25 features, pairwise pairwise mutual information was calculated for different lag steps. Taking the steel price index and the purchase price of a certain type of steel as an example, the mutual information value was calculated for each lag step from 0 to 60 days. It was found that the mutual information reached its maximum value of 0.68 at a lag of 15 days, exceeding the preset causality threshold of 0.3. Therefore, a directed edge was established from the steel price index to the purchase price of a certain type of steel, with the normalized mutual information value of 0.68 as the initial edge weight. Taking the exchange rate midpoint and the purchase price of imported materials as an example, it was found that the mutual information reached its maximum value of 0.55 at a lag of 7 days. A directed edge was established from the exchange rate midpoint to the purchase price of imported materials, with an initial edge weight of 0.55. Taking the technology reuse rate and the quota consumption of parts as an example, it was found that the mutual information reached its maximum value of 0.52 at a lag of 60 days. A directed edge was established from the technology reuse rate to the quota consumption of parts, with an initial edge weight of 0.52. Taking the intellectual property conversion rate of return and the technology reuse rate as examples, calculations show that the mutual information reaches its maximum value of 0.48 after a 90-day lag. A directed edge is then established from the intellectual property conversion rate of return to the technology reuse rate, with an initial edge weight of 0.48. Similarly, taking the correlation between the achievement and the R&D project and the technology reuse rate as examples, calculations show that the mutual information reaches its maximum value of 0.41 after a 30-day lag. A directed edge is then established from the correlation between the achievement and the R&D project to the technology reuse rate, with an initial edge weight of 0.41. Performing the above calculations on all feature pairs results in the establishment of 72 directed edges, forming a semantic constraint graph used to provide key causal paths for step S4.

[0042] Uncertainty quantification values ​​are calculated based on the semantic level to which each feature belongs, using the corresponding method. For structural features, the half-width of the statistical interval is used. Taking the hierarchical relationship of the bill of materials as an example, the half-width of the hierarchical relationship change interval in the past 12 data updates is 2.3%, and the uncertainty quantification value is 0.023. For dynamic features, the standard deviation of historical volatility is used. Taking the purchase price of a certain type of steel as an example, the standard deviation of its volatility over the past 12 weeks is calculated to be 8.7%, and the uncertainty quantification value is 0.087. Taking the time consumption of finishing processes as an example, the standard deviation of its deviation rate over the past 30 days is calculated to be 5.4%, and the uncertainty quantification value is 0.054. For externality features, the root mean square error of prediction is used. Taking the steel price index as an example, the root mean square error between its predicted and actual values ​​is calculated to be 5.2%, and the uncertainty quantification value is 0.052. Taking the exchange rate midpoint as an example, the root mean square error of its prediction is calculated to be 3.8%, and the uncertainty quantification value is 0.038. For outcome-based characteristics, the confidence interval width for predicted transformation revenue is used. Taking the technology reuse rate as an example, the confidence interval width for predicted transformation revenue is calculated to be 6.8%, with an uncertainty quantification value of 0.068. Taking the intellectual property transformation rate of return as an example, the confidence interval width is calculated to be 8.1%, with an uncertainty quantification value of 0.081. Taking the correlation between the outcome and the R&D project as an example, the confidence interval width is calculated to be 7.3%, with an uncertainty quantification value of 0.073.

[0043] The specific implementation of step S3 is as follows: Extract the out-degree and in-degree of each feature from the semantic constraint relationship graph and calculate the structural centrality. The out-degree of the steel price index node is 5, meaning this node points to 5 downstream features such as steel purchase price and component material cost; the in-degree is 0, indicating that no upstream feature points to it, and the calculated structural centrality value is 3.0. The out-degree of the exchange rate midpoint node is 4, and the in-degree is 0, so the calculated structural centrality value is 2.4. The out-degree of the steel purchase price node for a certain model is 3, and the in-degree is 2, so the calculated structural centrality value is 2.6. The out-degree of the technology achievement reuse rate node is 3, meaning this node points to 3 downstream features such as component quota consumption, process route optimization frequency, and number of R&D project initiations; the in-degree is 3, meaning it is affected by 3 upstream features such as intellectual property conversion rate of return, correlation between achievement and R&D project, and patent citation influence, so the calculated structural centrality value is 3.0. The out-degree of the equipment number node is 0, and the in-degree is 0, so the calculated structural centrality value is 0.

[0044] The confidence weights are calculated based on the quantified uncertainty values ​​of each feature, with a confidence decay coefficient set to 5. The quantified uncertainty value of the steel price index is 0.052, and the calculated confidence weight is 0.771. The quantified uncertainty value of the purchase price of a certain type of steel is 0.087, and the calculated confidence weight is 0.647. The quantified uncertainty value of the technology reuse rate is 0.068, and the calculated confidence weight is 0.712. The quantified uncertainty value of the equipment number is 0.015, and the calculated confidence weight is 0.928.

[0045] The initial fusion weights are obtained by multiplying the structural centrality of each feature by its confidence weight. Then, the initial fusion weights of all features are normalized to obtain the dynamic fusion weights. The initial fusion weight of the steel price index is 3.0 multiplied by 0.771, equaling 2.313. After normalization, its final dynamic fusion weight is 0.087. The initial fusion weight of the technology reuse rate is 3.0 multiplied by 0.712, equaling 2.136. After normalization, its final dynamic fusion weight is 0.080. The initial fusion weight of the exchange rate midpoint is 2.4 multiplied by 0.827, equaling 1.985. After normalization, its final dynamic fusion weight is 0.074. The initial fusion weight of the equipment number is 0 multiplied by 0.928, equaling 0. After normalization, its final dynamic fusion weight is 0. These 25 features are then weighted and fused according to the above dynamic fusion weights to generate the fused feature data.

[0046] When a new standardized dataset is generated at the beginning of each month, the dynamic fusion weights of each feature are automatically recalculated. For example, after a new dataset is added at the beginning of the following month, the historical volatility of the purchase price of a certain type of steel increases from 8.7% to 10.2%, its uncertainty quantification value increases to 0.102, the confidence weight decreases to 0.601, and the structural centrality changes due to the establishment of a new causal edge from exchange rate to imported material purchase price in the semantic constraint graph. After recalculation, its dynamic fusion weight is adjusted from 0.058 to 0.063. When step S6 updates the edge weight from technology reuse rate to component quota consumption in the semantic constraint graph based on the deviation analysis results, the change in edge weight causes the in-degree of the component quota consumption node to change from 3 to 4, and the structural centrality increases from 2.1 to 2.6, triggering a recalculation of its dynamic fusion weight from 0.042 to 0.055. Through the above two triggering conditions, the fusion weight is continuously and dynamically adjusted with data updates and causal structure evolution.

[0047] The specific implementation of step S4 is as follows: Using the unit cost of the product as the target node in the semantic constraint relationship graph constructed in step S2, trace back all causal paths pointing to this target node to extract three key causal paths. The first path is: the steel price index points to the steel purchase price, the steel purchase price points to the component material cost, and the component material cost points to the product unit cost. The second path is: the exchange rate midpoint points to the imported material purchase price, the imported material purchase price points to the component material cost, and the component material cost points to the product unit cost. The third path is: the technology reuse rate points to the component quota consumption, and the component quota consumption points to the product unit cost. These three key causal paths cover a total of 12 features.

[0048] A Long Short-Term Memory (LSTM) network prediction model is constructed. The input layer receives the time series of 25-dimensional feature data fused in step S3, with a 25-dimensional feature vector input at each time step. The causal path attention layer, located after the input layer, calculates attention weights for the 25-dimensional feature vector at each time step. The LSTM layer contains two LSTM layers, each with 64 hidden units, performing temporal modeling on the attention-masked feature sequence. The output layer is a fully connected layer that outputs the predicted value of the product's unit cost.

[0049] In the causal path attention layer, each of the 25 fused features is examined to determine whether it lies on one of the three key causal paths. The steel price index lies on the first key causal path, and its attention weight is retained, receiving an attention weight of 0.10 after softmax normalization. The exchange rate midpoint lies on the second key causal path, receiving an attention weight of 0.08 after retention. The technology reuse rate lies on the third key causal path, receiving an attention weight of 0.07 after retention. Nine other features, including steel purchase price, component material cost, imported material purchase price, and component quota consumption, also lie on key causal paths, each retaining its corresponding attention weight. Thirteen features, including equipment number, equipment depreciation data, crude oil price index, and the correlation between achievements and R&D projects, do not lie on any key causal path, and their attention weights are directly set to zero, remaining zero after normalization, and having no impact on subsequent calculations. After this processing, the prediction model's attention is forcibly constrained to 12 causally related features, ensuring that the prediction model makes predictions based solely on the fused features on the key causal paths.

[0050] The 25-dimensional feature sequence, constrained by attention masking, is input into a two-layer LSTM for time-series modeling. The model is trained using historical data from the past 36 months, with each month representing a time step, for a total of 36 time steps forming a training sample. A particle swarm optimization algorithm is used to search for the optimal hyperparameters, with a population size of 30, 100 iterations, and the fitness function being the root mean square error of prediction. After training, the optimal model parameters are obtained. Inputting the fused feature data from the current time step, the output layer generates a predicted unit cost of 495 yuan for the next month's product.

[0051] The specific implementation of step S5 is as follows: The actual unit cost of the company's products in the current month is 528 yuan, the dynamic forecast value generated in step S4 is 495 yuan, the deviation value is 33 yuan, and the deviation rate is 6.7%, exceeding the preset deviation threshold of 5%, triggering deviation attribution analysis. The integral gradient method is used as an interpretability analysis tool. Taking the input layer of the prediction model as the starting point and the output layer as the ending point, the integral gradient value of each of the 25 input features is calculated as the contribution of each input feature to the prediction result. The calculation results show that the contribution of the steel price index is positive (12.5), pushing the predicted value higher. The contribution of the purchase price of a certain type of steel is positive (9.8). The contribution of the exchange rate midpoint is positive (6.3). The contribution of the intellectual property conversion rate of return is positive (2.9). The contribution of the technology reuse rate is negative (5.2), pushing the predicted value lower. The contribution of the equipment number is 0.1, with almost no impact.

[0052] Using the causal path in the semantic constraint graph constructed in step S2 as the propagation path, the contribution of each input feature is propagated layer by layer along the causal path to the target node of product unit cost. The contribution of the steel price index (12.5) propagates along the causal edge to the steel purchase price. After adding its own contribution (9.8) and the propagated contribution of 12.5 from the steel price index, the cumulative contribution of the steel purchase price is 22.3, which continues to propagate along the causal edge to the component material cost. The contribution of the exchange rate midpoint (6.3) propagates along the causal edge to the imported material purchase price. The imported material purchase price itself has a contribution of 4.2, and its cumulative contribution is 10.5, which propagates to the component material cost. The component material cost, after aggregating the contributions from both the steel purchase price and the imported material purchase price paths, has a cumulative contribution of 32.8, which propagates to the product unit cost. The contribution of the technology reuse rate (-5.2) propagates along the causal edge to the component quota consumption. The component quota consumption itself has a contribution of 2.3, and its cumulative contribution is -2.9, which propagates to the product unit cost.

[0053] After propagation to the unit cost of the product, the cumulative contribution is summarized according to the semantic level. The cumulative contribution of the externality level is 18.8, the cumulative contribution of the dynamic level is 14.0, the cumulative contribution of the outcome level is 8.1, and the cumulative contribution of the structural level is 3.5. The feature with the highest cumulative contribution in the externality level is the steel price index; the feature with the highest cumulative contribution in the dynamic level is the steel purchase price; the feature with the highest cumulative contribution in the outcome level is the technology reuse rate; and the feature with the highest cumulative contribution in the structural level is the bill of materials hierarchy.

[0054] Comparing the cumulative contributions of each semantic level, the externality level has the largest cumulative contribution at 18.8, followed by the dynamic level at 14.0, the outcome level at 8.1, and the structural level at 3.5. Since the cumulative contributions of the externality and dynamic levels are significantly higher than other levels, the deviation type is labeled as a mixed-class deviation. A deviation feedback signal is generated, containing a deviation value of 33 yuan, a mixed-class deviation type, and contribution characteristics of steel price index, steel purchase price, technology achievement reuse rate, and bill of materials hierarchy for each semantic level.

[0055] The specific implementation of step S6 is as follows: Construct a deviation pattern knowledge base. After three months of system operation, the company has accumulated 87 historical records. Each record includes the deviation type, deviation magnitude, contribution characteristics at each semantic level, a snapshot of the semantic constraint relationship graph, the optimization strategy adopted, and evaluation indicators of the optimization effect. Record number 52 represents a deviation event in March 2025. The deviation type is mixed, with a deviation rate of 7.2%. The externality level contribution characteristic is the steel price index, the dynamic level contribution characteristic is the steel purchase price, and the outcome level contribution characteristic is the technology achievement reuse rate. The optimization strategy adopted was to increase the fusion weight of the steel price index and the steel purchase price and shorten the supply chain data update cycle. After optimization, the deviation rate decreased from 7.2% to 4.1%.

[0056] The current deviation feedback signal was semantically matched with each of the 87 historical records in the knowledge base. The deviation type was identified as mixed, with a deviation rate of 6.7%. The externality level contribution feature was the steel price index, the dynamic level contribution feature was the steel procurement price, and the outcome level contribution feature was the technology reuse rate. This deviation, along with record number 52, is also a mixed deviation type. The deviation amplitude ranges of 7.2% and 6.7% both fall within the 5% to 10% range. The contribution features across the three semantic levels—externality, dynamics, and outcome—are completely consistent, indicating a successful match.

[0057] The optimization strategy corresponding to historical record number 52 is used as the initial reference strategy. This includes increasing the fusion weight of the steel price index and steel procurement price features, and shortening the update cycle of supply chain data. Adaptive fine-tuning is performed based on the latest edge weights in the current semantic constraint graph. The edge weight from the technology reuse rate to the component quota consumption in the current graph has increased from 0.52 to 0.61, indicating that the causal effect of technology reuse on cost reduction is enhanced. Therefore, based on the initial reference strategy, the fusion weight of the technology reuse rate feature is further increased to 0.075.

[0058] The strategy for reusing research results was adjusted and implemented concurrently. Five reusable technological achievements were pushed to the R&D team through the technology innovation achievement sharing service system, of which three were adopted and applied to the current R&D project, increasing the technology achievement reuse rate from 45% to 57%. The strategy for patent and trademark search and identification was also adjusted and implemented concurrently. The accuracy and scope of patent searches were improved for three key technologies involved in the current R&D project, identifying one potential intellectual property conflict risk and adjusting the R&D plan in a timely manner.

[0059] After implementing the optimization strategy, an effectiveness evaluation was conducted. The supply chain data update cycle was shortened from 7 days to 3 days, and the R&D cost prediction deviation rate decreased from 6.7% to 3.5%. The deviation feedback signal, the implemented optimization strategy, and the optimization effect evaluation indicators were stored as the 88th record in the deviation pattern knowledge base.

[0060] If no matching historical deviation pattern is found in the knowledge base for the current deviation feedback signal, the knowledge discovery process is triggered. The time lag mutual information calculation in step S2 is re-executed for the contribution features determined by deviation attribution in step S5, and the corresponding causal edge weights are updated. After the edge weights are updated, the edge weight update condition in step S3 is triggered, and the dynamic fusion weights of each feature are automatically recalculated.

[0061] After six consecutive cycles of closed-loop iterative optimization, the average deviation rate of cost prediction for this R&D project decreased from the initial 7.2% to 3.1%, the frequency of cost overruns decreased by approximately 42%, and the reuse rate of technological achievements cumulatively increased from 45% to 73%, a cumulative increase of 28 percentage points. The semantic constraint graph expanded from the initial 72 directed edges to 76 directed edges after two knowledge discovery processes. The causal structure continuously evolved with the actual deviation feedback of R&D activities, and the multi-source data fusion strategy achieved dynamic self-optimization.

[0062] In step S1, the unified identifier includes adding a timestamp and a data source system identifier to each data record. The technological innovation achievement data includes the technology achievement reuse rate and achievement transformation rate of return. The intellectual property data includes the patent legal status and patent citation relationship.

[0063] It should be noted that unified identification refers to mapping fields from various data sources to standard fields according to a unified data dictionary, and adding a timestamp and data source system identifier to each data record.

[0064] Technological innovation achievement data refers to data from the technological innovation achievement sharing service system and the technology transfer and promotion management system, including the technology reuse rate and achievement transfer rate. The technology reuse rate is the ratio of the frequency of existing technological achievements being cited in subsequent R&D projects to the total number of achievements. The achievement transfer rate is the ratio of the economic benefits brought by achievement transfer to the corresponding R&D investment.

[0065] Intellectual property data refers to data from intelligent systems for precise retrieval and identification of patents and trademarks, including patent legal status and patent citation relationships. Patent legal status refers to the current legal stage of a patent, such as authorization, examination, or expiration. Patent citation relationships refer to the relationships between prior patents cited by this patent and subsequent patents that cite it.

[0066] Step S2 further includes the following sub-steps: S2-1, semantic hierarchical annotation of features from each data source is performed, dividing the features into four semantic levels: structural features, dynamic features, external features, and outcome features, and defining directed constraint relationships between each semantic level to eliminate semantic heterogeneity between different data sources. S2-2, calculate the time lag mutual information for any two features under different lag step sizes, take the lag direction corresponding to the maximum mutual information value as the causal direction, if the mutual information value is greater than the preset causal threshold, establish a directed edge from cause to effect to eliminate the temporal heterogeneity between different data sources, construct a semantic constraint relationship graph, and the semantic constraint relationship graph is used to provide key causal paths; S2-3, calculate the uncertainty quantification value according to the semantic level to which each feature belongs, so as to eliminate the quality heterogeneity between different data sources. Among them, structural features are calculated based on statistical intervals, dynamic features are calculated based on historical volatility, external features are calculated based on prediction error, and outcome features are calculated based on the confidence level of conversion benefit prediction.

[0067] Furthermore, in sub-step S2-1, structural features include the hierarchical relationship of the bill of materials and the standard process route; dynamic features include fluctuations in procurement prices and changes in labor consumption; external features include market material price indices and exchange rate fluctuations; and outcome features include the reuse rate of technological achievements, intellectual property transformation benefits, and the correlation between achievements and R&D projects.

[0068] Furthermore, in sub-step S2-1, the directed constraint relationships include: structural features constrain the value range of dynamic features, externality features affect the evolution direction of dynamic features, dynamic features affect the output efficiency of outcome features, and outcome features feed back into the adjustment of structural features.

[0069] It should be noted that structural characteristics refer to relatively stable basic attributes in R&D activities, including the bill of materials hierarchy and standard process routes. The bill of materials hierarchy refers to the parent-child hierarchy and subordination relationship between components in the product structure, and the standard process route refers to the standard sequence and quota hours of each process in the product manufacturing process.

[0070] Dynamic characteristics refer to cost drivers with time-series evolution characteristics, including fluctuations in purchase prices and changes in labor consumption. Fluctuations in purchase prices refer to the range of changes in the unit price of the same component in different purchase batches, while changes in labor consumption refer to the degree of deviation between the actual labor hours and the standard labor hours of the same process in different periods.

[0071] Externalities refer to environmental variables from outside the R&D system, including market material price indices and exchange rate fluctuations. Market material price indices refer to the monthly price changes of major raw materials in the open market, while exchange rate fluctuations refer to the changes in the exchange rate between foreign currency and RMB used for corporate settlement.

[0072] Outcome characteristics refer to indicators that reflect the value of R&D output and the ability to provide technological innovation services. These include the reuse rate of technological achievements, the revenue from intellectual property transformation, and the correlation between achievements and R&D projects. The reuse rate of technological achievements refers to the ratio of the frequency of existing technological achievements being cited in subsequent R&D projects to the total number of achievements. The revenue from intellectual property transformation refers to the economic income generated by patents or trademarks through licensing, transfer, etc. The correlation between achievements and R&D projects refers to the degree of matching between technological achievements and their source R&D projects and the quantification of their contribution.

[0073] Directed constraint relationships refer to the causal constraints and transmission relationships between semantic levels, including the range of values ​​of structural features constrained by dynamic features, the evolution direction of dynamic features influenced by external features, the output efficiency of outcome features influenced by dynamic features, and the adjustment of structural features by outcome features in turn.

[0074] Different lag steps refer to the preset time offset range in the time lag mutual information calculation. This range is determined based on the maximum observable delay of cost transmission in R&D activities, for example, set to 0 to 90 days. Within this range, all lag steps are traversed in days to calculate the mutual information value under each lag step.

[0075] Time lag mutual information refers to the calculation of mutual information of time series of any two features under different lag steps. The lag direction corresponding to the maximum mutual information value is taken as the causal direction. If the mutual information value is greater than the preset causal threshold, a directed edge from cause to effect is established, and the normalized mutual information value is used as the initial edge weight.

[0076] The preset causal threshold is the minimum mutual information value used to determine whether there is a causal relationship between two features. When the calculated time lag mutual information value is greater than the threshold, a causal relationship is identified and a directed edge is established. When the mutual information value is less than or equal to the threshold, a causal relationship is identified and no directed edge is established. The threshold is determined based on the mutual information distribution characteristics of historical data, and is usually taken as the upper quantile of the mutual information value distribution.

[0077] This application systematically addresses the fusion barriers of multi-source heterogeneous data across three dimensions: semantics, temporal sequence, and quality, through semantic hierarchical annotation, time-lag mutual information causal discovery, and differentiated uncertainty quantification. Features are divided into four semantic levels: structural, dynamic, external, and outcome-based, and directed constraints are defined between these levels, enabling isolated features to establish structured relationships based on the inherent logic of R&D cost evolution. The time-lag mutual information method is used to automatically identify the causal direction and propagation delay between features, constructing a semantic constraint graph with causal relationships as directed edges. Uncertainty quantification values ​​are calculated using different methods at each semantic level, assigning lower weights to features with lower data quality during fusion.

[0078] Step S3 further includes the following sub-steps: S3-1, extract the out-degree and in-degree of each feature from the semantic constraint relationship graph, and calculate the structure centrality; S3-2, Calculate the confidence weight based on the uncertainty quantification value of each feature; S3-3, combining structural centrality and confidence weights, generates dynamic fusion weights to weight and fuse features from various data sources; S3-4 triggers the re-execution of S3-1 to S3-3 when a new standardized dataset is generated or the edge weights of the semantic constraint graph are updated, updating the dynamic fusion weights of each feature.

[0079] It should be noted that the confidence weight refers to the weight coefficient that reflects the credibility of the feature data, calculated based on the uncertainty quantification value of each feature. The higher the uncertainty quantification value, the lower the confidence weight.

[0080] The confidence decay coefficient is a preset parameter used to control the influence of uncertainty quantification values ​​on confidence weights. The larger the coefficient, the stronger the decay effect of uncertainty on confidence, that is, the higher the degree to which features with poor data quality are suppressed in the fusion. The coefficient is determined according to the overall distribution level of uncertainty quantification values ​​of each feature.

[0081] This application addresses the problem of relying solely on a single statistical indicator for weight allocation in multi-source data fusion by comprehensively generating dynamic fusion weights from two independent dimensions: causal topology and data credibility. For each feature, out-degree and in-degree are extracted from the semantic constraint graph to calculate structural centrality, reflecting topological importance. Simultaneously, confidence weights, reflecting data credibility, are calculated based on uncertainty quantification values. These two dimensions are then combined to generate dynamic fusion weights. When a new standardized dataset is generated or the edge weights of the semantic constraint graph are updated, the dynamic fusion weights are automatically recalculated, allowing the fusion strategy to continuously adjust with data updates and the evolution of the causal structure.

[0082] Please see Figure 2 A flowchart of the prediction model provided in an embodiment of the present invention.

[0083] Step S4 further includes the following sub-steps: S4-1, Extract all causal paths pointing to the cost prediction target node from the semantic constraint relationship graph constructed in step S2, and use them as key causal paths; S4-2, Construct a prediction model for a long short-term memory network. The prediction model includes an input layer, a causal path attention layer, a long short-term memory layer, and an output layer. S4-3, In the causal path attention layer, the attention weights of the features located on the key causal path in the features fused in step S3 are retained, and the attention weights of the features not on the key causal path are reset to zero, so that the prediction model makes predictions only based on the fused features on the key causal path. S4-4 inputs the attention-constrained feature sequence into the long short-term memory layer for time-series modeling, and the output layer generates dynamic prediction results of R&D costs.

[0084] It should be noted that the causal path attention layer refers to the intermediate layer located between the input layer and the long short-term memory layer in the long short-term memory network prediction model. It applies causal consistency weighting to the input features, so that the model prioritizes features that have a direct causal relationship with the prediction target when making predictions.

[0085] Attention weight preservation refers to maintaining the normal weight calculation of features located on key causal paths in the attention mechanism, so that they can participate in subsequent time series modeling and prediction.

[0086] Resetting attention weights to zero means setting the attention weights of features that are not on the critical causal path to zero, so that they have no effect in subsequent long short-term memory layers.

[0087] This application addresses the problem of deep learning prediction models being susceptible to spurious correlation features and having poorly interpretable results by introducing causal paths from the semantic constraint graph as hard masks for attention. All key causal paths pointing to the cost prediction target are extracted from the semantic constraint graph. In the attention layer, the weights of features not on causal paths are forcibly set to zero, ensuring the model makes predictions based solely on causally validated features. This hard constraint mechanism injects causal prior knowledge into the deep learning model, making prediction results traceable to specific causal paths while eliminating interference from spurious correlation features.

[0088] Step S5 further includes the following sub-steps: S5-1, When the deviation between the actual cost data and the dynamic prediction result generated in step S4 exceeds the preset deviation threshold, the interpretability analysis tool is used to obtain the contribution of each input feature in the prediction model of step S4 to the prediction result. S5-2, using the causal path in the semantic constraint relationship graph constructed in step S2 as the propagation path, the contribution of each feature is propagated layer by layer along the causal path to the cost prediction target node. Based on the cumulative contribution of each semantic level during the propagation process, the feature with the highest cumulative contribution is determined as the contributing feature of that semantic level. S5-3, compare the cumulative contribution of each semantic level, and take the deviation type corresponding to the semantic level with the largest cumulative contribution as the deviation type of this time. When the cumulative contribution of the outcome feature is the largest, it is marked as an outcome-type deviation, and a deviation feedback signal containing the deviation value, deviation type and contribution features of each semantic level is generated.

[0089] It should be noted that interpretability analysis tools refer to algorithms used to quantify the contribution of each input feature in a prediction model to the prediction result. Specifically, the integral gradient method can be used to measure the influence of each feature on the final predicted value by calculating the gradient path integral from the input feature to the predicted output.

[0090] Contribution refers to the quantitative impact of each input feature on the prediction result, calculated by interpretability analysis tools. A positive value indicates that the feature increases the predicted value, a negative value indicates that the feature decreases the predicted value, and the larger the absolute value, the stronger the impact.

[0091] Propagation along the causal path refers to using the causal path in the semantic constraint relationship graph as the propagation channel to pass and accumulate the contribution of the feature level from the cause node to the effect node layer by layer, so that the contribution of the upstream feature converges to the downstream feature along the direction of the causal edge, and finally converges to the cost prediction target node.

[0092] The cumulative contribution of semantic levels refers to the cumulative contribution value of all features in each of the four semantic levels—structural, dynamic, external, and outcome—after the contribution is propagated layer by layer along the causal path to the cost prediction target node.

[0093] The contribution feature of a semantic level refers to the feature with the highest cumulative contribution in each semantic level, representing the specific factor that contributes the most to the bias in that level.

[0094] The preset deviation threshold is a critical value used to determine whether the deviation between the actual cost and the predicted result is significant. When the deviation rate exceeds the threshold, deviation attribution analysis is triggered; when the deviation rate does not exceed the threshold, attribution analysis is not triggered. This threshold is determined based on the historical cost fluctuation range of the R&D project and the business's tolerance for cost deviation.

[0095] The deviation type refers to the classification of deviation causes determined by comparing the cumulative contribution of each semantic level. If the cumulative contribution of structural features is the largest, it is marked as the first type of deviation; if the cumulative contribution of dynamic features is the largest, it is marked as the second type of deviation; if the cumulative contribution of externality features is the largest, it is marked as the third type of deviation; if the cumulative contribution of outcome features is the largest, it is marked as outcome-type deviation; if the cumulative contribution of multiple semantic levels is close, it is marked as a mixed type of deviation.

[0096] This application addresses the problem that existing bias analysis methods can only locate individual features and cannot trace the transmission chain or determine the causal level by propagating feature contributions layer by layer along the causal path in the semantic constraint graph and classifying them according to semantic level. Interpretable analysis tools are used to obtain the contribution of each input feature to the prediction result, which is then propagated layer by layer along the causal path to the cost prediction target node. The cumulative contribution is summarized according to the semantic level, and the contribution feature at each level is determined. The type of bias is determined by comparing the cumulative contribution at each level. This attribution method can distinguish whether the bias is caused by external market changes, internal cost fluctuations, or insufficient technology transfer, providing structured bias feedback signals for subsequent optimization strategy matching.

[0097] Step S6 further includes the following sub-steps: S6-1, Construct a deviation pattern knowledge base to store historical deviation patterns and their corresponding optimization strategies; S6-2, based on the semantic hierarchical matching results between the current deviation feedback signal and the historical deviation pattern, invoke the corresponding optimization strategy or trigger the knowledge discovery process; S6-3 If the corresponding optimization strategy is invoked, the weights of the causal edges related to the contribution features determined by the bias attribution in step S5 in the semantic constraint relationship graph are adjusted according to the execution effect of the optimization strategy. Otherwise, the time lag mutual information calculation in step S2 is re-executed for the contribution features determined by the bias attribution in step S5, and the corresponding causal edge weights are updated.

[0098] Furthermore, in sub-step S6-3, the optimization strategy includes adjusting the results reuse strategy and adjusting the patent and trademark search and identification strategy; adjusting the results reuse strategy involves pushing reusable technological results through the technology innovation results sharing service system; adjusting the patent and trademark search and identification strategy involves improving the accuracy and scope of patent and trademark searches when deviation analysis indicates the existence of intellectual property conflict risks.

[0099] It should be noted that calling the corresponding optimization strategy means that when a historical deviation pattern of semantic level matching is retrieved, the optimization strategy corresponding to the historical pattern is used as the initial reference strategy, and then executed after adaptive fine-tuning in combination with the current semantic constraint relationship graph.

[0100] Adjusting the weights of causal edges in the semantic constraint graph refers to increasing or decreasing the weights of causal edges related to the contribution characteristics determined by the bias attribution in step S5, based on the execution effect of the optimization strategy.

[0101] Adjusting the results reuse strategy means that when deviation analysis shows that R&D cost overruns are related to low technology reuse rates, existing reusable technological achievements are pushed to the R&D team through the technology innovation achievement sharing service system in order to improve the technology reuse rate and reduce redundant R&D costs.

[0102] Adjusting the patent and trademark search and identification strategy refers to adjusting the parameters of the strategy to improve search accuracy and scope when deviation analysis indicates that the research and development direction may conflict with existing intellectual property rights, thereby avoiding redundant research and development and infringement costs caused by intellectual property conflicts.

[0103] This application addresses the problems of deviations merely serving as error records and failing to be transformed into reusable knowledge, as well as the lack of adaptive learning capabilities when encountering unseen deviation patterns, by constructing a deviation pattern knowledge base and performing historical deviation pattern retrieval, matching, and closed-loop updates based on semantic hierarchy. After each deviation attribution, the generated deviation feedback signal is stored in the knowledge base to form a historical record. When a new deviation occurs, a weighted matching process is performed based on the semantic hierarchy of deviation type, deviation amplitude range, and contribution features to retrieve similar historical deviation patterns. Upon successful matching, historical optimization strategies are reused and adaptively fine-tuned according to the current semantic constraint relationship graph before execution. Optimization strategies include improving the reuse rate of technological achievements and mitigating intellectual property conflict risks. Upon failed matching, a knowledge discovery process is triggered, recalculating the time-lag mutual information of contribution features to update the causal edge weights. After the edge weights are updated, a recalculation of the fusion weights is automatically triggered, thereby affecting the attention mask of the prediction model and subsequent deviation attribution results. This allows the semantic constraint relationship graph to continuously evolve with actual deviation feedback, achieving an evolution from passive response to proactive self-optimization.

[0104] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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 this application, and should all be included within the protection scope of this application.

Claims

1. A method for dynamic optimization of R&D costs based on multi-source data fusion, characterized in that, Includes the following steps: Step S1: Obtain multi-source heterogeneous R&D data, perform unified identification and preprocessing, and generate a standardized dataset. The multi-source heterogeneous R&D data includes technological innovation achievement data and intellectual property data. Step S2: Semantic hierarchical annotation of features from each data source is performed, time lag mutual information method is used to identify causal relationships between features, semantic constraint relationship graph with causal relationships as directed edges is constructed, and uncertainty quantification value of each feature is calculated. Step S3: Based on the structural centrality and uncertainty quantification values ​​of each feature in the semantic constraint relationship graph, generate dynamic fusion weights, perform weighted fusion of features from each data source, and recalculate the dynamic fusion weights of each feature when the standardized dataset or semantic constraint relationship graph is updated. Step S4: Extract key causal paths pointing to the cost prediction target from the semantic constraint relationship graph, construct a long short-term memory network prediction model with the features on the key causal paths as attention masks, and generate dynamic prediction results of R&D costs. Step S5: Calculate the deviation between the actual cost data and the prediction results. Based on the interpretability analysis results of the Long Short-Term Memory Network prediction model, perform deviation attribution analysis along the causal path and semantic hierarchy in the semantic constraint relationship graph to generate a deviation feedback signal. Step S6: Based on the semantic hierarchical matching results of the current deviation feedback signal and the historical deviation pattern, call the corresponding optimization strategy or trigger the knowledge discovery process, and update the edge weights and dynamic fusion weights of the semantic constraint relationship graph.

2. The method for dynamic optimization of R&D costs based on multi-source data fusion according to claim 1, characterized in that: In step S1, the unified identifier includes a timestamp and a data source system identifier for each data record; the technological innovation achievement data includes the technology achievement reuse rate and achievement transformation rate of return; and the intellectual property data includes the patent legal status and patent citation relationship.

3. The method for dynamic optimization of R&D costs based on multi-source data fusion according to claim 1, characterized in that: Step S2 further includes the following sub-steps: S2-1, semantic hierarchical annotation of features from each data source is performed, dividing the features into four semantic levels: structural features, dynamic features, external features, and outcome features, and defining directed constraint relationships between each semantic level to eliminate semantic heterogeneity between different data sources. S2-2, calculate the time lag mutual information for any two features under different lag step sizes, take the lag direction corresponding to the maximum value of mutual information as the causal direction, if the mutual information value is greater than the preset causal threshold, establish a directed edge from cause to effect to eliminate the temporal heterogeneity between different data sources, and construct a semantic constraint relationship graph, which is used to provide key causal paths; S2-3, calculate the uncertainty quantification value according to the semantic level to which each feature belongs, so as to eliminate the quality heterogeneity between different data sources. Among them, structural features are calculated based on statistical intervals, dynamic features are calculated based on historical volatility, external features are calculated based on prediction error, and outcome features are calculated based on the confidence level of conversion benefit prediction.

4. The method for dynamic optimization of R&D costs based on multi-source data fusion according to claim 3, characterized in that, In sub-step S2-1, the structural features include the material bill of materials hierarchy and standard process routes; the dynamic features include fluctuations in purchase prices and changes in labor consumption; the external features include market material price indices and exchange rate fluctuations; and the outcome features include the reuse rate of technological achievements, intellectual property transformation revenue, and the correlation between achievements and R&D projects.

5. The method for dynamic optimization of R&D costs based on multi-source data fusion according to claim 3, characterized in that, In sub-step S2-1, the directed constraint relationship includes: structural features constrain the value range of dynamic features, externality features affect the evolution direction of dynamic features, dynamic features affect the output efficiency of outcome features, and outcome features feed back into the adjustment of structural features.

6. The method for dynamic optimization of R&D costs based on multi-source data fusion according to claim 1, characterized in that: Step S3 further includes the following sub-steps: S3-1, extract the out-degree and in-degree of each feature from the semantic constraint relationship graph, and calculate the structure centrality; S3-2, Calculate the confidence weight based on the uncertainty quantification value of each feature; S3-3, combining structural centrality and confidence weights, generates dynamic fusion weights to weight and fuse features from various data sources; S3-4 triggers the re-execution of S3-1 to S3-3 when a new standardized dataset is generated or the edge weights of the semantic constraint graph are updated, updating the dynamic fusion weights of each feature.

7. The method for dynamic optimization of R&D costs based on multi-source data fusion according to claim 1, characterized in that: Step S4 further includes the following sub-steps: S4-1, Extract all causal paths pointing to the cost prediction target node from the semantic constraint relationship graph constructed in step S2, and use them as key causal paths; S4-2, Construct a long short-term memory network prediction model, the prediction model including an input layer, a causal path attention layer, a long short-term memory layer and an output layer; S4-3, In the causal path attention layer, the attention weights of the features located on the key causal path in the features fused in step S3 are retained, and the attention weights of the features not on the key causal path are reset to zero, so that the prediction model makes predictions only based on the fused features on the key causal path. S4-4 inputs the attention-constrained feature sequence into the long short-term memory layer for time-series modeling, and the output layer generates dynamic prediction results of R&D costs.

8. The method for dynamic optimization of R&D costs based on multi-source data fusion according to claim 1, characterized in that: Step S5 further includes the following sub-steps: S5-1, When the deviation between the actual cost data and the dynamic prediction result generated in step S4 exceeds the preset deviation threshold, the interpretability analysis tool is used to obtain the contribution of each input feature in the prediction model of step S4 to the prediction result. S5-2, using the causal path in the semantic constraint relationship graph constructed in step S2 as the propagation path, the contribution of each feature is propagated layer by layer along the causal path to the cost prediction target node. Based on the cumulative contribution of each semantic level during the propagation process, the feature with the highest cumulative contribution is determined as the contributing feature of that semantic level. S5-3, compare the cumulative contribution of each semantic level, and take the deviation type corresponding to the semantic level with the largest cumulative contribution as the deviation type of this time. When the cumulative contribution of the outcome feature is the largest, it is marked as an outcome-type deviation, and a deviation feedback signal containing the deviation value, deviation type and contribution features of each semantic level is generated.

9. The method for dynamic optimization of R&D costs based on multi-source data fusion according to claim 1, characterized in that: Step S6 further includes the following sub-steps: S6-1, Construct a deviation pattern knowledge base to store historical deviation patterns and their corresponding optimization strategies; S6-2, based on the semantic hierarchical matching results between the current deviation feedback signal and the historical deviation pattern, invoke the corresponding optimization strategy or trigger the knowledge discovery process; S6-3, if the corresponding optimization strategy is invoked, the weights of the causal edges related to the contribution features determined by the bias attribution in step S5 in the semantic constraint relationship graph are adjusted according to the execution effect of the optimization strategy; otherwise, the time lag mutual information calculation in step S2 is re-executed for the contribution features determined by the bias attribution in step S5, and the corresponding causal edge weights are updated.

10. The method for dynamic optimization of R&D costs based on multi-source data fusion according to claim 9, characterized in that, In sub-step S6-3, the optimization strategy includes adjusting the results reuse strategy and adjusting the patent and trademark retrieval and identification strategy; the adjusted results reuse strategy is to push reusable technological results through the technology innovation results sharing service system; the adjusted patent and trademark retrieval and identification strategy is to improve the accuracy and scope of patent and trademark retrieval when the deviation analysis indicates that there is a risk of intellectual property conflict.

Images