Vibrating compaction forming energy transfer efficiency evaluation method

Abstract

The invention relates to the technical field of data processing, in particular to a vibration compaction forming energy transfer efficiency evaluation method, which comprises the following steps: firstly, integrating multi-source data, and generating a semantic enhancement data set through a self-adaptive semantic annotation algorithm; then, node relation extraction is carried out based on the semantic enhancement data set, and a knowledge graph model is constructed and dynamically maintained by using a first graph algorithm so as to structurally represent nodes and relations in the compaction process; based on the knowledge graph model, a parameter influence result is evaluated by a second graph algorithm, and a risk prediction result is generated in combination with a time sequence prediction algorithm; then, a reward function is defined according to a parameter influence result and a risk prediction result, and an optimal vibration compaction parameter combination is determined; and finally, generating a parameter recommendation report and collecting user feedback to iteratively optimize the knowledge graph. According to the method, the accuracy of vibration compaction parameter selection can be improved, and the energy transmission efficiency and the engineering quality are optimized.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a method for evaluating energy transfer efficiency of vibration compaction molding. Background Art

[0002] The final quality and energy efficiency of vibration compaction projects are highly nonlinearly coupled with the physical and mechanical properties of the fillers used, the performance parameters of the compaction equipment, and the refined control strategy of the compaction process. In order to ensure the long-term stability and performance of, for example, high-grade highway subgrades, high-speed railway subgrades, or backfill areas of critical infrastructure, the engineering community has put forward increasingly stringent requirements for the matching of filler gradation, maximum particle size, moisture content, and compaction equipment selection and operating parameters. Traditional vibration compaction parameter selection relies heavily on engineer experience, local standards, or verification based on a small number of field tests. This approach often fails to fully account for variable working conditions and material differences, and can easily lead to low compaction energy utilization and insufficient compaction uniformity, or problems such as local overpressure and underpressure.

[0003] Reinforcement learning, as a machine learning method that can learn optimal strategies through interaction with the environment, provides a new approach to solving the above problems. By constructing a reinforcement learning model suitable for the field of vibration compaction, scattered geological data, material constitutive model parameters, equipment performance curves, construction specification requirements, compaction plans and effect data of historical projects, and even a large number of CAE simulation data sets can be used as part of the training data or environmental state for structured utilization. However, simply introducing a reinforcement learning framework does not mean that strategy optimization has been achieved. The core technical challenge currently faced is the lack of effective means to intelligently recommend a compaction parameter sequence that adapts to the current working conditions through reinforcement learning, achieve strategy optimization, and dynamically optimize the energy distribution of the entire compaction process to avoid problems such as "jump vibration", excessive energy consumption, reduced equipment life, or excessive material crushing caused by improper selection of excitation energy. To this end, a method for evaluating the energy transfer efficiency of vibration compaction molding is proposed. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for evaluating the energy transfer efficiency of vibration compaction molding, aiming to improve the accuracy of vibration compaction parameter selection and optimize energy transfer efficiency and engineering quality. This method first integrates multi-source data and generates a semantically enhanced dataset using an adaptive semantic annotation algorithm based on machine learning and graph clustering. Subsequently, a GNN algorithm is applied to extract node relationships, and a first graph algorithm is used to construct and dynamically maintain a knowledge graph model to structure the node relationships of the compaction process. Based on this knowledge graph model, a second graph algorithm is used to evaluate parameter impact results, and a time series prediction algorithm is combined to generate risk prediction results. Furthermore, a reward function is defined based on the parameter impact results and the risk prediction results, and a deep reinforcement learning algorithm is used to determine the optimal vibration compaction parameter combination to achieve strategy optimization. Finally, a parameter recommendation report is generated and user feedback is collected to iteratively optimize the knowledge graph.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A method for evaluating energy transfer efficiency of vibration compaction molding, comprising:

[0007] Integrate multi-source heterogeneous vibration compaction data and generate a semantically enhanced dataset through an adaptive semantic annotation algorithm based on machine learning and graph clustering algorithms;

[0008] Based on the semantically enhanced dataset, a domain-enhanced GNN algorithm is applied to extract node relationships, and incremental updates are performed using the first graph algorithm to construct and dynamically maintain a knowledge graph model; the knowledge graph model is used to structurally represent the nodes and relationships in the vibration compaction process;

[0009] Based on the knowledge graph model, a second graph algorithm is used to evaluate parameter impact results; and based on the node embedding features of the knowledge graph model, a time series prediction algorithm is used to generate a risk prediction result; a reward function is defined according to the parameter impact results and the risk prediction results, and an optimal vibration compaction parameter combination is determined using a deep reinforcement learning algorithm;

[0010] A parameter recommendation report is generated based on the optimal vibration compaction parameter combination and the knowledge graph path, and user feedback is collected to iteratively optimize the knowledge graph model.

[0011] Furthermore, the implementation process of the adaptive semantic annotation algorithm includes:

[0012] generating a field embedding vector for each raw data field in the multi-source heterogeneous vibrocompaction data using a machine learning algorithm fine-tuned on a target engineering domain dataset;

[0013] Inputting the field embedding vector into a graph clustering algorithm, the graph clustering algorithm clusters the original data fields according to the similarity of the embedding vectors, and establishing dynamic labeling rules for each cluster;

[0014] The dynamic annotation rules are applied to map the original data fields to standard concepts and attributes in a predefined vibration compaction domain ontology, and the semantic labels formed by the mapping are consistency checked by referring to the vibration compaction domain ontology to generate the semantically enhanced dataset.

[0015] Furthermore, the process of extracting node relationships includes:

[0016] Converting the semantically enhanced dataset into an initial graph structure, wherein the initial graph structure is used to represent the identified key data elements in the semantically enhanced dataset with nodes, and to represent the established semantic classification information and preliminary structural links between the key data elements with relationships;

[0017] Using a GNN algorithm pre-trained on a target engineering domain dataset, distributed representation learning is performed on the nodes in the initial graph structure to identify key engineering entities; relationship classification is performed based on the key engineering entities to form structured knowledge triples;

[0018] The validity of the structured knowledge triple is verified with a predefined vibration compaction domain ontology, and the verified structured knowledge triple is output.

[0019] Furthermore, the process of performing incremental updates using the first graph algorithm to build and dynamically maintain the knowledge graph model includes:

[0020] When newly added semantically enhanced data is received, generating a temporal embedding representation for the entity interaction events in the newly added semantically enhanced data using a first graph algorithm;

[0021] Combined with a lightweight graph neural network, newly added structured knowledge triples are extracted from the temporal embedding representation to form a local newly added graph;

[0022] Through the graph difference detection algorithm, the changed part of the local newly added graph is identified to be integrated into the knowledge graph model, and an update operation is performed on the corresponding changed part of the knowledge graph model; the update operation includes adding, modifying and deleting nodes and relationships, and performing validity verification based on the predefined vibration compaction domain ontology.

[0023] Furthermore, the process of using the second graph algorithm to evaluate the parameter impact results includes: extracting a dynamic context subgraph from the knowledge graph model using a graph segmentation algorithm based on the current compaction conditions and candidate vibration compaction parameters; inputting the dynamic context subgraph into the pre-trained second graph algorithm; the second graph algorithm is used to capture the spatiotemporal dependencies and node interaction patterns in the dynamic context subgraph, infer and output parameter impact results, and the parameter impact results are quantitative representations of the candidate vibration compaction parameters on predefined compaction effect indicators.

[0024] Furthermore, the process of generating risk prediction results includes: extracting a dynamic context subgraph from the knowledge graph model using a graph segmentation algorithm based on the current compaction conditions and candidate vibration compaction parameters; applying a node embedding algorithm to generate a node embedding vector representation for the nodes in the dynamic context subgraph, and inputting the node embedding vector representation as a node embedding feature into a pre-trained time series prediction algorithm to output the risk prediction results corresponding to the candidate vibration compaction parameters.

[0025] Furthermore, the process of determining the optimal vibration compaction parameter combination using the deep reinforcement learning algorithm includes:

[0026] Constructing a deep reinforcement learning environment, including: defining an agent state, including current compaction working condition characteristics, material properties, equipment status, parameter impact results, and risk prediction results extracted from the knowledge graph model; defining an agent action, including selecting a next set of vibration compaction parameters; defining an agent reward function, including: an expected compaction quality improvement evaluated based on the parameter impact results, expected energy consumption and equipment loss inferred from the knowledge graph model, and an amount of risk imposed for exceeding a safety threshold based on the risk prediction results;

[0027] Based on the deep reinforcement learning environment, a deep reinforcement learning algorithm is used to train a parameter selection strategy that can maximize the long-term cumulative expected reward. The output of the parameter selection strategy is the optimal vibration compaction parameter combination.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] 1. This invention integrates heterogeneous vibratory compaction data from multiple sources, including computer-aided engineering (CAE) data, through an adaptive semantic annotation algorithm. Based on this, a domain-enhanced GNN algorithm and advanced first-graph algorithm technology are employed to construct and dynamically maintain a structured vibratory compaction domain knowledge graph model. This provides a unified, comprehensive knowledge and data foundation capable of mapping operating condition changes in real time for subsequent energy transfer efficiency assessment, parameter impact analysis, risk prediction, and intelligent optimization decision-making.

[0030] 2. This invention accurately assesses parameter impacts through the construction of a dynamic knowledge graph model, combined with a second graph algorithm. Furthermore, by integrating the knowledge graph's node embedding features with a time-series prediction algorithm, it effectively predicts operational risks. This multi-dimensional, intelligent analysis and prediction capability reveals the dynamic impact of various parameters on compaction performance indicators during complex vibration compaction processes, proactively identifying potential risks. This not only enhances the understanding of the vibration compaction process and the accuracy of predictions, but also provides a solid foundation for refined and adaptive process control.

[0031] 3. Based on a comprehensive assessment of parameter impacts and risks, the present invention integrates a reward function defined by knowledge graph model attributes with a deep reinforcement learning algorithm that integrates risk constraints to achieve a multi-objective optimization of vibration compaction process parameter combinations. This method can accurately determine the optimal parameter combination that ensures compaction quality and avoids operational risks. It uses the knowledge graph path to generate explanatory parameter recommendation reports to assist engineers in making scientific decisions. In addition, a user feedback mechanism is introduced to continuously iterate and optimize the knowledge graph model, forming a learning and improvement closed loop. This effectively improves the energy transfer efficiency, construction quality, and equipment utilization of vibration compaction molding, optimizing energy transfer efficiency and project quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 The present invention provides a flow chart of a method for evaluating energy transfer efficiency of vibration compaction molding;

[0033] Figure 2 A flowchart of the entity relationship extraction process is provided for the present invention;

[0034] Figure 3 The present invention provides a flow chart for determining the optimal vibration compaction parameter combination using a deep reinforcement learning algorithm. DETAILED DESCRIPTION

[0035] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0036] See also Figures 1 to 3 The present invention provides a method for evaluating the energy transfer efficiency of vibration compaction molding, and the technical solution is as follows:

[0037] Example 1:

[0038] Vibratory compaction is a critical process in the construction of infrastructure such as roads and foundations. Its compaction quality and energy efficiency directly impact the economic viability and durability of the project. However, this process involves equipment characteristics, complex multi-source material data, and the dynamic interaction between numerous process parameters, making optimal control and energy efficiency evaluation extremely challenging. Traditional methods often rely on empirical or isolated analysis, making it difficult to integrate heterogeneous data for comprehensive and dynamic parameter optimization and efficiency evaluation, often resulting in low energy efficiency and fluctuating quality.

[0039] In order to overcome the above limitations and improve the intelligence level and energy utilization efficiency of vibration compaction, this embodiment describes in detail a vibration compaction energy transfer efficiency evaluation method, taking the vibration compaction process of a highway subgrade as an example. Figure 1 As shown, specifically including:

[0040] Multi-source heterogeneous vibration compaction data are integrated and a semantically enhanced dataset is generated through an adaptive semantic annotation algorithm based on machine learning algorithm and graph clustering algorithm.

[0041] Among them, multi-source heterogeneous vibration compaction data includes but is not limited to: computer-aided engineering (CAE) simulation analysis output files, such as CSV format or text format output containing information such as stress field, strain field, density distribution, etc.; real-time data collected by sensors at the compaction site, including but not limited to vibration frequency, amplitude, rolling speed, GPS positioning information, etc., recorded in the form of logs or streaming data; database records of historical engineering projects, which contain construction process parameters, material mechanical properties indicators, final compaction quality data, etc.; standardized material physical and mechanical properties test reports, such as test data in PDF or electronic spreadsheet format.

[0042] Furthermore, multi-source heterogeneous vibration compaction data suffers from format diversity and heterogeneous semantic expressions, which affects the efficiency and accuracy of data fusion and semantic unification. To address these challenges, the implementation process of the adaptive semantic annotation algorithm includes:

[0043] Firstly, a field embedding vector is generated for each raw data field in the multi-source heterogeneous vibrocompaction data using a machine learning algorithm fine-tuned on a target engineering domain dataset.

[0044] The term "raw data field" refers to the original name or identifier for a specific piece of information in the data source, such as the column name "Freq_Hz" in a CSV file or the key name "VibrationFrequency" in a sensor log. In this example, a Transformer-based BERT model is pre-trained on relevant highway corpus for these fields to extract context-sensitive semantic vectors. Each field is then converted into a field embedding vector.

[0045] The field embedding vectors are then fed into a graph clustering algorithm (such as the Louvain community discovery algorithm or K-Means clustering), which clusters the original data fields based on the similarity of the embedding vectors (such as cosine similarity or Euclidean distance). Even if the original names are different, such as "Freq_Hz" and "VibrationFrequency," they will be assigned to the same cluster because their embedding vectors are close in vector space.

[0046] Subsequently, for each formed field cluster, dynamic annotation rules are generated automatically or with minimal human intervention. These rules define how to map all raw data fields within the cluster to a unified semantic meaning, laying the foundation for subsequent automated semantic alignment and unification.

[0047] Finally, the dynamic annotation rules are applied to map the raw data fields to the standard concepts and attributes in the predefined vibration compaction domain ontology. The vibration compaction domain ontology is a shared knowledge framework pre-built by the vibration compaction profession, with standard concepts such as "compaction machinery type", "soil material", "vibration parameters" and other key categories or things. For the standard concept of "vibration parameter type", its attributes may include "parameter name", "parameter value", "standard unit" and so on. When processing "raw data fields" such as "Freq_Hz", it corresponds to the attribute "operating frequency" under the standard concept of "vibration parameter type" in the ontology, and its unit should be Hertz and the data type should be numerical.

[0048] During the mapping process, consistency checks are performed on the semantically labeled results based on the attribute constraints defined in the ontology (including data type restrictions, legal value ranges, hierarchical relationships, and mutually exclusive relationships). Only when all labeled results meet consistency and logical requirements will they be included in the final output semantically enhanced dataset.

[0049] The resulting dataset, after verification, contains both raw data and standardized semantic labels, providing high-quality input for subsequent knowledge graph model construction. Field embeddings are generated using a fine-tuned pre-trained model, and dynamic annotation rules are established in conjunction with a graph clustering algorithm, enabling adaptive annotation of multi-source heterogeneous data. Compared to template matching methods, this approach better handles data diversity, non-standardized naming, and complex terminology, improving annotation automation and accuracy. This facilitates the construction of an accurate and reliable vibration compaction knowledge graph model, enhancing the intelligence of evaluation and optimization methods.

[0050] Based on the semantically enhanced dataset, a domain-enhanced GNN algorithm is applied to extract node relationships, and the first graph algorithm is used to perform incremental updates to build and dynamically maintain a knowledge graph model; the knowledge graph model is used to structuredly represent the nodes and relationships in the vibration compaction process.

[0051] Furthermore, if Figure 2 As shown in Figure 2, the process of entity relationship extraction includes:

[0052] First, the semantic enhancement dataset is converted into an initial graph structure. The initial graph structure includes:

[0053] Used to represent the identified key data elements in the semantically enhanced dataset with nodes. The nodes can represent: a specific model of compaction equipment (such as "20-ton single-steel wheel vibratory roller of model XXX"); construction parameter values ​​(such as "vibration frequency 35Hz", "rolling speed 2.5km / h").

[0054] Relationships are used to represent established semantic classification information and preliminary structural links between key data elements. Relationships can represent: preliminary semantic relationships formed based on existing semantic labels in semantically enhanced data, such as the "XP303 road roller" node pointing to a class node in the ontology through a relationship; and preliminary structural links within the data source itself, such as the inherent data associations between geometric models, material constitutive parameters, and stress-strain calculation results in CAE simulation data.

[0055] Then, a GNN algorithm (such as GraphSAGE or GAT) that has been pre-trained on a dataset of the target engineering field (such as road engineering and geotechnical engineering) is used to process the initial graph structure. This includes: performing distributed representation learning on the nodes in the initial graph structure to obtain their node feature vectors, which integrate the attributes of the node itself and the semantic information of its structural neighborhood. Based on these learned node feature vectors, key engineering entities (such as roadbed structural units, compaction equipment, and key construction parameters, etc.) that are of great significance to the analysis of the vibration compaction process are further identified; then, relationship classification is performed based on the key engineering entities, and their relationship types are mapped to semantic relationship types predefined in the domain ontology. For example, the roadbed level contains filler types, and the compaction equipment uses vibration parameters to form structured knowledge triples, each of which is in the form of [head entity, relationship type, tail entity];

[0056] Finally, the structured knowledge triples are verified for validity against the predefined vibration compaction domain ontology, which defines constraint rules for each entity type and relationship type. For example, a certain type of relationship can only be established between entities of a specific type. Only when the triples meet the above-mentioned semantic rationality and structural consistency requirements, the verified structured knowledge triples are output for subsequent incremental updates or initial construction of the knowledge graph model.

[0057] By converting semantically enhanced datasets into an initial graph structure, the domain-enhanced GNN algorithm model is used to perform entity representation learning and relationship classification, accurately extracting deep entities and complex semantic relationships from semi-structured or structured data. Pre-training and domain ontology verification further ensure the accuracy of the extraction results, providing technical support for building a high-quality knowledge graph model for the vibration compaction domain.

[0058] Furthermore, the process of performing incremental updates using the first graph algorithm to build and dynamically maintain the knowledge graph model includes:

[0059] First, when new semantically enhanced data is received, a first graph algorithm, the Temporal Graph Network (TGN) model, is used to generate a temporal embedding representation for the entity interaction events in the newly added semantically enhanced data. For example, the event of a specific type of compaction equipment, a "vibratory roller," performing its third pass of compaction on the A2 layer of roadbed fill on the highway section from K10+200 to K10+300, will be encoded as an embedding representation with temporal context. This representation not only reflects the core features of the event itself, but also incorporates the temporal information of the event, enhancing the model's understanding of the dynamic evolution process.

[0060] Subsequently, a lightweight graph neural network (GNN) is combined to extract new structured knowledge triples from the temporal embedding representation. These triples represent new pieces of knowledge extracted from the latest semantically enhanced data and need to be integrated into the main knowledge graph model. Logically, they form a local additional graph, a small temporary graph structure containing the latest dynamic information.

[0061] Finally, a graph difference detection algorithm (e.g., Graph Edit Distance) is used to identify the changes that integrate the newly added graph into the knowledge graph model, focusing on subgraph regions that are highly relevant to the topic of the newly added information. To ensure that the newly added triples can be seamlessly integrated into the existing graph structure, an update operation is performed on the corresponding changed portions of the knowledge graph model. This update operation includes adding, modifying, and deleting nodes and relationships. A validation check is also performed based on the predefined vibration compaction domain ontology to ensure that all newly integrated information is semantically consistent with the domain knowledge specification and that the overall structure is reasonable and usable.

[0062] By using the first-graph algorithm to process newly added dynamic data, leveraging a lightweight graph neural network to extract new knowledge, and implementing targeted incremental updates using a graph difference detection algorithm, this approach achieves efficient dynamic maintenance of the knowledge graph model in the vibratory compaction domain. Only the modified parts of the model are updated, reducing computational overhead and improving update efficiency. Furthermore, continuous domain ontology verification during the update process ensures the model's quality and consistency during dynamic evolution, providing a reliable and up-to-date knowledge foundation for higher-level applications that rely on the model for real-time analysis and decision-making.

[0063] Based on the knowledge graph model, a second graph algorithm is used to evaluate the parameter impact results; and based on the node embedding features of the knowledge graph model, a time series prediction algorithm is used to generate a risk prediction result; a reward function is defined according to the parameter impact results and the risk prediction results, and a deep reinforcement learning algorithm is used to determine the optimal vibration compaction parameter combination.

[0064] Furthermore, a second graph algorithm, also a Temporal Graph Network (TGN) model, is used to predict the future impact of specific inputs (candidate parameters). The process of evaluating the impact of parameters includes:

[0065] First, obtain current compaction condition information, including the type of material used in the construction area, the material's moisture content, the number of compaction passes completed, and the ambient temperature on site. This condition data can be extracted from existing knowledge graph models or input by on-site users.

[0066] Under the premise of knowing the current compaction working conditions, multiple candidate vibration compaction parameter combinations are constructed. The parameter content usually includes different vibration frequencies, amplitudes and rolling speeds. Subsequently, information related to the above working conditions and candidate parameters is extracted from the knowledge graph model. In order to extract effective knowledge structure fragments, a graph segmentation algorithm (such as the Louvain community discovery algorithm) is used to divide the knowledge graph model and extract a dynamic context subgraph from it. This dynamic context subgraph not only contains entity information directly related to the current working conditions and candidate parameters, such as the model and attributes of specific materials and compaction equipment, but also includes the evolutionary state and historical interaction process of these entities in the time dimension, reflecting their behavioral characteristics that change over time.

[0067] This dynamic contextual subgraph is then fed into a pre-trained second-graph algorithm. This network is capable of capturing the temporal and spatial dependencies and interactions between nodes and relationships within the graph, identifying the complex mechanisms by which parameter changes influence the material compaction process. The network then infers and outputs parameter impact results, which are quantitative representations of the candidate vibration compaction parameters' impact on predefined compaction performance metrics under the current compaction conditions.

[0068] These compaction effect indicators include but are not limited to: the density increment of the next pass, the uniformity coefficient of stress distribution, the effective utilization rate of energy, and the time required to reach the target density. By calculating the impact of parameters on the results, they provide key reference basis for the subsequent parameter selection and optimization process.

[0069] A dynamic context subgraph is constructed based on the working conditions, and the second graph algorithm is combined to deeply infer the parameter utility to achieve dynamic prediction of the consequences of vibration compaction parameters. It can make full use of the context semantics and time evolution information in the knowledge graph to enhance the prediction fidelity and applicability, and provide high-quality input and state transition prediction basis for vibration compaction parameter optimization based on reinforcement learning and other means, thereby improving the accuracy of vibration compaction parameter selection.

[0070] Furthermore, the process of generating risk prediction results includes:

[0071] First, relevant information is obtained based on the current compaction conditions and candidate vibration compaction parameters. Current compaction conditions typically include the type of material in the construction area, moisture content, number of compaction passes, and ambient temperature. Candidate vibration compaction parameters include different combinations of vibration frequency, amplitude, and rolling speed. Based on this input information, graph structure information relevant to risk prediction is extracted from a dynamically maintained knowledge graph model.

[0072] To achieve this goal, a graph segmentation algorithm is used to extract a dynamic contextual subgraph from the knowledge graph model. This subgraph contains entities and their attributes related to risk events, such as typical operating conditions that have historically caused risks, material sensitivity parameters, and equipment limit state records, providing a structured foundation for subsequent risk analysis.

[0073] Next, a node embedding algorithm is applied to generate node embedding vector representations for the nodes in the dynamic context subgraph. The node embedding algorithm used, such as Node2Vec or GraphSAGE, maps each node in the subgraph to a low-dimensional real vector. By analyzing the node's neighborhood structure, connectivity patterns, and semantic information, this vector embedding effectively captures the node's position in the graph and its potential connections with other entities, thereby forming discriminative node embedding features.

[0074] Subsequently, the node embedding vector representation is used as a node embedding feature and input into the pre-trained time series prediction algorithm to output the risk prediction results corresponding to the candidate vibration compaction parameters. The model can be a classification model or a regression model, and specific types include but are not limited to logistic regression, support vector machine or gradient boosting tree (such as XGBoost). The model aims to predict potential risks in the compaction construction process and can identify operational risks such as the probability or level of "jump vibration" risk, "particle crushing" degree, and "equipment load abnormality". After receiving the input node embedding feature, the model performs the inference process and outputs the risk prediction results corresponding to each set of candidate vibration compaction parameters. The output risk prediction results can be probability values, risk scores or risk level labels, reflecting the potential operational risk levels of different parameter combinations under the current working conditions.

[0075] By combining the semantic expression capabilities of the knowledge graph, the structural perception capabilities of the node embedding algorithm, and the predictive capabilities of the time series prediction algorithm, it is possible to accurately identify potential operational risks under complex compaction conditions, providing support for risk avoidance and intelligent optimization decisions of subsequent vibration compaction parameters, thereby improving the accuracy of vibration compaction parameter selection and optimizing energy transfer efficiency and project quality.

[0076] Further, if Figure 3 As shown in Figure 2, the process of determining the optimal vibration compaction parameter combination using the deep reinforcement learning (DRL) algorithm includes:

[0077] First, we build a deep reinforcement learning environment. The construction of this environment includes the definition of the following core elements:

[0078] The agent state is defined as the environmental information observed by the agent at each decision moment. This state contains data inputs from multiple dimensions, including: current compaction condition characteristics extracted from the knowledge graph model, such as the construction area number, current compaction layer thickness, number of completed passes, ambient temperature and humidity, etc.; material properties, such as soil type, particle gradation, moisture content, target dry density, etc.; equipment status, such as roller model, current vibration frequency, amplitude, driving speed, etc.; parameter impact results, i.e., the predicted value of compaction quality (such as density and uniformity) based on the aforementioned parameters; and risk prediction results, i.e., the probability or risk level of the candidate parameter combination causing operational risks such as vibration hopping, particle breakage, and equipment overload under the current working conditions.

[0079] These multi-dimensional information together form the basis for the DRL agent's decision-making.

[0080] Define agent actions. Define the actions the agent can perform, primarily selecting the next set of vibration compaction parameters. This parameter set can include control variables such as vibration frequency, amplitude, rolling speed, and whether to continue compaction. The action space can be set to discrete (selecting from a preset parameter set) or continuous (dynamically adjusting parameter values ​​within a specified range).

[0081] Define the agent reward function and construct a reward function to guide learning behavior. This function comprehensively considers factors such as compaction quality, energy consumption, equipment loss, and operational risk. Specifically, it provides positive rewards for improving compaction quality, such as increased density and uniformity; imposes moderate negative penalties for energy consumption and equipment loss based on the estimated fuel or electricity consumption of the selected parameters, as well as equipment wear and life reduction; and imposes negative penalties for risk factors, such as behaviors that may cause operational risk indicators (such as bounce probability and particle breakage rate) to exceed safety thresholds. The reward function can use a weighted sum or other function form to combine these indicators, ultimately generating a scalar reward value that reflects the overall quality of the current action.

[0082] Once the environment is constructed, specific deep reinforcement learning algorithms, such as Proximal Policy Optimization (PPO), Deep Q Network (DQN), and their improved variants, are applied for policy training. By executing numerous "state-action-reward" loops within the simulation environment, the agent continuously updates its policy function, gradually learning a policy model that outputs the optimal combination of vibration compaction parameters under various compaction conditions.

[0083] By modeling the vibration compaction parameter optimization problem as a deep reinforcement learning task and comprehensively utilizing the parameter impact assessment and risk prediction information provided by the knowledge graph model in a constructed multi-dimensional fusion environment, a self-adaptive and globally optimal vibration compaction parameter determination mechanism is implemented. This enables optimal strategy selection in a multi-objective trade-off, improving the accuracy of vibration compaction parameter selection and optimizing energy transfer efficiency and project quality.

[0084] A parameter recommendation report is generated based on the optimal vibration compaction parameter combination and the knowledge graph path. At the same time, the system collects feedback information such as engineers' adoption of the recommended solutions and actual construction results. This feedback information is used to continuously iterate and optimize the content of the knowledge graph model, forming a closed-loop learning and improvement mechanism.

[0085] By constructing and dynamically maintaining a knowledge graph model for vibration compaction that integrates multi-source heterogeneous data, including computer-aided engineering analysis, and applying advanced algorithms such as neural networks, first-graph algorithms, and deep reinforcement learning, the limitations of traditional vibration compaction methods in data integration, process understanding, parameter optimization, and risk control have been overcome. This not only enables accurate assessment and prediction of the dynamic impact of compaction parameters and potential construction risks, but also automatically finds the optimal multi-objective process parameter combination that takes into account energy transfer efficiency, compaction quality, and operational safety. Ultimately, by providing explainable decision support and establishing a continuous learning feedback mechanism, the accuracy of vibration compaction parameter selection is improved, and energy transfer efficiency and project quality are optimized.

[0086] Example 2:

[0087] Based on Example 1, Example 2 further demonstrates the application efficiency of the present invention in engineering scenarios, ensuring that high energy transfer efficiency is still achieved under complex working conditions. A method for evaluating the energy transfer efficiency of vibration compaction molding includes:

[0088] Integrate multi-source heterogeneous vibration compaction data and generate a semantically enhanced dataset through an adaptive semantic annotation algorithm based on machine learning and graph clustering algorithms;

[0089] Based on the semantically enhanced dataset, a domain-enhanced GNN algorithm is applied to extract node relationships, and incremental updates are performed using the first graph algorithm to construct and dynamically maintain a knowledge graph model; the knowledge graph model is used to structurally represent the nodes and relationships in the vibration compaction process;

[0090] Based on the knowledge graph model, a second graph algorithm is used to evaluate parameter impact results; and based on the node embedding features of the knowledge graph model, a time series prediction algorithm is used to generate a risk prediction result; a reward function is defined according to the parameter impact results and the risk prediction results, and an optimal vibration compaction parameter combination is determined using a deep reinforcement learning algorithm;

[0091] A parameter recommendation report is generated based on the optimal vibration compaction parameter combination and the knowledge graph path, and user feedback is collected to iteratively optimize the knowledge graph model.

[0092] For comparison purposes, this example considers a common roadbed fill type (Group A graded crushed stone) and sets clear compaction targets (average compaction of 95% and a coefficient of variation of compaction of less than 4%). The inventive method is compared with two baseline methods:

[0093] Baseline Method 1: Traditional empirical parameter method. Fixed vibration frequency, amplitude, and rolling speed are set based on typical construction manuals and operator experience.

[0094] Baseline method 2: Simplified model-based optimization algorithm (standard genetic algorithm): This method is based on a simplified filler compaction model (without complex knowledge graph and real-time fusion of multi-source data) and uses a genetic algorithm to optimize the compaction parameters offline.

[0095] All methods were run under the same simulation conditions and data input to ensure a fair comparison. As shown in Table 1, the performance indicators and overall efficiency of the vibration compaction process are presented. Among them, the average energy transfer efficiency directly reflects the effectiveness of energy utilization. The method of the present invention is expected to achieve a significantly higher percentage, which shows that through intelligent parameter optimization, it can more effectively use the energy output of the compaction equipment to improve the material density, thereby reducing energy waste; at the same time, the final average compaction degree is used to measure the compaction quality. The present invention is not only easier to stably achieve and moderately exceed the target value of the engineering design; the compaction degree standard deviation reflects the compaction uniformity. The smaller the value, the better the uniformity. The present invention can improve the uniformity of the compaction operation and avoid engineering quality defects caused by local over- or under-pressure, thereby comprehensively ensuring the overall performance and long-term stability of the compaction structure. In addition, thanks to the higher energy transfer efficiency and more sophisticated parameter control strategy, the present invention also shows advantages in the average operation time to achieve the target compaction degree, which can complete the compaction task more quickly, thereby improving overall construction efficiency and economic benefits.

[0096] Table 1 Comparison of compaction performance and efficiency

[0097] Performance indicators Baseline Method 1 Baseline Method 2 Method of the present invention Average energy transfer efficiency (%) 62 70 83 Final average compaction degree (%) 94.2 95.1 95.8 Compaction standard deviation 1.5 1.1 0.7 <![CDATA[Average operation duration (hours / 100m) to reach the target compaction degree 2 )]]> 5.5 4.8 3.9

[0098] In order to further verify the effect of the present invention, a specific working condition was simulated, in which the roadbed material type was set to medium-sized graded crushed stone, the initial moisture content was 8%, the compaction layer thickness was 30 cm, and the target compaction degree was 95%. As shown in Table 2, a comparison of compaction performance and efficiency is given. It can be seen that the method of the present invention, through its intelligent optimization decision-making, can recommend slightly different frequencies and amplitudes for the current working conditions (materials, targets, etc., which have been integrated into the knowledge graph model), and through a more optimal combination of passes and speeds, while achieving a better compaction effect, the expected total energy consumption is also lower than that of other baseline methods.

[0099] Table 2 Comparison of compaction parameters and performance results of different methods under specific working conditions

[0100]

[0101] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for evaluating energy transfer efficiency of vibration compaction molding, characterized in that: include: Integrate multi-source heterogeneous vibration compaction data and generate a semantically enhanced dataset through an adaptive semantic annotation algorithm based on machine learning and graph clustering algorithms; Based on the semantically enhanced dataset, a domain-enhanced GNN algorithm is applied to extract node relationships, and an incremental update is performed using the first graph algorithm to construct and dynamically maintain a knowledge graph model; The knowledge graph model is used to structurally represent nodes and relationships in the vibration compaction process; Based on the knowledge graph model, a second graph algorithm is used to evaluate parameter impact results; and based on the node embedding features of the knowledge graph model, a time series prediction algorithm is used to generate a risk prediction result; a reward function is defined according to the parameter impact results and the risk prediction results, and an optimal vibration compaction parameter combination is determined using a deep reinforcement learning algorithm; A parameter recommendation report is generated based on the optimal vibration compaction parameter combination and the knowledge graph path, and user feedback is collected to iteratively optimize the knowledge graph model.

2. A vibration compaction molding energy transfer efficiency evaluation method according to claim 1, characterized in that: The implementation process of the adaptive semantic annotation algorithm includes: generating a field embedding vector for each raw data field in the multi-source heterogeneous vibrocompaction data using a machine learning algorithm fine-tuned on a target engineering domain dataset; Inputting the field embedding vector into a graph clustering algorithm, the graph clustering algorithm clusters the original data fields according to the similarity of the embedding vectors, and establishing dynamic labeling rules for each cluster; The dynamic annotation rules are applied to map the original data fields to standard concepts and attributes in a predefined vibration compaction domain ontology, and the semantic labels formed by the mapping are checked for consistency through the vibration compaction domain ontology to generate the semantically enhanced dataset.

3. The method for evaluating energy transfer efficiency of vibration compaction molding according to claim 1, characterized in that: The process of node relationship extraction includes: Converting the semantically enhanced dataset into an initial graph structure, wherein the initial graph structure is used to represent the identified key data elements in the semantically enhanced dataset with nodes, and to represent the established semantic classification information and preliminary structural links between the key data elements with relationships; Using a GNN algorithm pre-trained on a target engineering domain dataset, distributed representation learning is performed on the nodes in the initial graph structure to identify key engineering entities; relationship classification is performed based on the key engineering entities to form structured knowledge triples; The validity of the structured knowledge triple is verified with a predefined vibration compaction domain ontology, and the verified structured knowledge triple is output.

4. The method for evaluating energy transfer efficiency of vibration compaction molding according to claim 1, characterized in that: The process of using the first graph algorithm to perform incremental updates to build and dynamically maintain the knowledge graph model includes: When newly added semantically enhanced data is received, generating a temporal embedding representation for the entity interaction events in the newly added semantically enhanced data using a first graph algorithm; Combined with a lightweight graph neural network, newly added structured knowledge triples are extracted from the temporal embedding representation to form a local newly added graph; Through the graph difference detection algorithm, the changed part of the local newly added graph is identified to be integrated into the knowledge graph model, and an update operation is performed on the corresponding changed part of the knowledge graph model; the update operation includes adding, modifying and deleting nodes and relationships, and performing validity verification based on the predefined vibration compaction domain ontology.

5. The method for evaluating energy transfer efficiency of vibration compaction molding according to claim 1, characterized in that: The process of using the second graph algorithm to evaluate the parameter impact results includes: extracting a dynamic context subgraph from the knowledge graph model using a graph segmentation algorithm based on the current compaction conditions and candidate vibration compaction parameters; inputting the dynamic context subgraph into the pre-trained second graph algorithm; the second graph algorithm is used to capture the spatiotemporal dependencies and node interaction patterns in the dynamic context subgraph, infer and output parameter impact results, and the parameter impact results are a quantitative representation of the candidate vibration compaction parameters on the predefined compaction effect indicators.

6. The method for evaluating energy transfer efficiency of vibration compaction molding according to claim 1, characterized in that: The process of generating risk prediction results includes: extracting a dynamic context subgraph from the knowledge graph model using a graph segmentation algorithm based on the current compaction conditions and candidate vibration compaction parameters; applying a node embedding algorithm to generate a node embedding vector representation for the nodes in the dynamic context subgraph, and inputting the node embedding vector representation as a node embedding feature into a pre-trained time series prediction algorithm to output the risk prediction results corresponding to the candidate vibration compaction parameters.

7. The method for evaluating energy transfer efficiency of vibration compaction molding according to claim 1, characterized in that: The process of using deep reinforcement learning algorithms to determine the optimal vibration compaction parameter combination includes: Constructing a deep reinforcement learning environment, including: defining an agent state, including current compaction working condition characteristics, material properties, equipment status, parameter impact results, and risk prediction results extracted from the knowledge graph model; defining an agent action, including selecting a next set of vibration compaction parameters; defining an agent reward function, including: an expected compaction quality improvement evaluated based on the parameter impact results, expected energy consumption and equipment loss inferred from the knowledge graph model, and an amount of risk imposed for exceeding a safety threshold based on the risk prediction results; Based on the deep reinforcement learning environment, a deep reinforcement learning algorithm is used to train a parameter selection strategy that can maximize the long-term cumulative expected reward. The output of the parameter selection strategy is the optimal vibration compaction parameter combination.

Images