Digital path dynamic generation method and system

Abstract

The invention discloses a digital path dynamic generation method and system, and the method comprises the steps: obtaining multi-dimensional heterogeneous data, carrying out the preprocessing and feature extraction of the multi-dimensional heterogeneous data, and obtaining a preliminarily converted feature matrix; obtaining a standardized feature vector set through nonlinear manifold representation construction and dimension reduction processing; outputting an initial path template and a corresponding key node sequence; converting a path optimization problem into a transmission problem from source distribution to target distribution through an optimal transmission theory in combination with user individual characteristics and resource constraint conditions, and performing solution optimization to obtain a personalized path scheme; in combination with real-time monitoring data and feedback information, a dynamic adjustment decision model is constructed through a Bayesian network modeling inter-node conditional probability relationship and a reinforcement learning algorithm, path node adjustment operation is automatically executed, and a dynamically updated personalized path is output.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method and system for dynamically generating digital paths, which can be applied to complex scenarios requiring strict process control and significant individual differences, such as precision manufacturing, healthcare, and complex project management, to achieve optimized resource allocation and refined process management. Background Technology

[0002] In the field of modern digital management, path optimization and the construction of personalized decision-making processes have become key technologies for improving system efficiency and accuracy. These technologies are typically applied to complex scenarios requiring strict process control and significant individual differences, such as precision manufacturing, healthcare, and complex project management, to achieve optimized resource allocation and refined process management.

[0003] Current mainstream technologies typically employ rule-based static template matching methods, using pre-defined decision trees or flowcharts to guide operational processes. For example, some systems use case-based reasoning (CBR) to construct standard process templates based on historical success cases, and then select the most similar template for application based on user-input feature parameters. Another approach uses a statistical model-based scheme, establishing a correlation model between parameters and the optimal path by mining historical data.

[0004] However, the above technologies have significant shortcomings: First, existing systems struggle to effectively handle high-dimensional heterogeneous data, especially when the input features are high-dimensional and diverse in type, making it difficult to accurately capture the complex correlations between features; second, existing technologies lack true dynamic adjustment capabilities, failing to automatically adjust subsequent paths based on real-time data feedback during the process, resulting in the entire process failing to achieve optimal results if the initial selection is not ideal; furthermore, existing systems exhibit low optimization efficiency when dealing with the temporal relationships and resource constraints of path nodes, making it difficult to find the globally optimal solution under multi-objective constraints. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method and system for dynamic generation of digital paths. This method involves introducing Hamiltonian-Poisson integrator technology to process multidimensional heterogeneous data, applying a stochastic neural network to reconstruct a multidimensional random field model, employing a convex relaxation method for high-dimensional optimal transmission to perform personalized path optimization, and combining Bayesian networks and reinforcement learning algorithms to construct a dynamic adjustment model, thereby achieving real-time dynamic adjustment of path nodes. Finally, Monte Carlo simulation technology is used to verify the feasibility of the dynamic path. This invention solves the problems of existing technologies being unable to effectively process high-dimensional heterogeneous data, lacking true dynamic adjustment capabilities, and exhibiting low optimization efficiency.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for dynamically generating digital paths includes: acquiring multidimensional heterogeneous data; preprocessing and extracting features from the multidimensional heterogeneous data to obtain a preliminary transformed feature matrix; based on the preliminary transformed feature matrix, constructing and reducing the dimensionality through nonlinear manifold representation to obtain a standardized feature vector set; based on the standardized feature vector set and a pre-built knowledge graph and standard path template library, constructing a mapping relationship between features and path templates through a random neural network and training the model to output an initial path template and corresponding key node sequence; according to the initial path template and corresponding key node sequence, combined with user individual characteristics and resource constraints, transforming the path optimization problem into a transmission problem from source distribution to target distribution through optimal transmission theory and solving and optimizing it to obtain a personalized path scheme; based on the personalized path scheme, combined with real-time monitoring data and feedback information, modeling the conditional probability relationship between nodes through a Bayesian network and constructing a dynamic adjustment decision model through reinforcement learning algorithm, automatically executing path node adjustment operations, and outputting a dynamically updated personalized path; based on the dynamically updated personalized path, applying Monte Carlo methods... Carlo simulation technology generates simulation scenarios and performs path simulations. Adjustments are made through statistical analysis and risk assessment, and the final executable optimized path and corresponding decision support information are output.

[0008] The beneficial effects of this invention are:

[0009] 1. This invention introduces Hamiltonian-Poisson integrator technology to process multidimensional heterogeneous data. By solving a system of nonlinear partial differential equations, it achieves a unified representation of heterogeneous data while maintaining the data topology, thus solving the efficiency and accuracy problems of traditional feature engineering when processing high-dimensional heterogeneous data.

[0010] 2. This invention applies a stochastic neural network to reconstruct a multidimensional random field model, transforming the deterministic path selection problem into a probabilistic inference problem, effectively capturing the complex correlations and uncertainties between features, and improving the accuracy and robustness of path template matching. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A flowchart of the method for dynamically generating digital paths provided in an embodiment of the present invention;

[0013] Figure 2A flowchart of the multidimensional heterogeneous data preprocessing and feature extraction steps provided in the embodiments of the present invention. Detailed Implementation

[0014] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0015] like Figure 1 As shown, this embodiment provides a method for dynamically generating digital paths, the method including:

[0016] S1. Obtain multidimensional heterogeneous data, preprocess and extract features from the multidimensional heterogeneous data to obtain a preliminary transformed feature matrix;

[0017] S2. Based on the feature matrix of the preliminary transformation, a standardized feature vector set is obtained through nonlinear manifold representation construction and dimensionality reduction processing;

[0018] S3. Based on the standardized feature vector set and the pre-built knowledge graph and standard path template library, a mapping relationship between features and path templates is constructed through a random neural network and the model is trained to output the initial path template and the corresponding key node sequence;

[0019] S4. Based on the initial path template and the corresponding key node sequence, combined with user individual characteristics and resource constraints, the path optimization problem is transformed into a transmission problem from source distribution to target distribution through optimal transmission theory and then solved and optimized to obtain a personalized path scheme.

[0020] S5. Based on the personalized path scheme, combined with real-time monitoring data and feedback information, a dynamic adjustment decision model is constructed by modeling the conditional probability relationship between nodes through a Bayesian network and combining it with a reinforcement learning algorithm. The path node adjustment operation is automatically executed, and a dynamically updated personalized path is output.

[0021] S6. Based on the dynamically updated personalized path, apply Monte Carlo simulation technology to generate a simulation scenario and run the path simulation. Adjustments are made through statistical analysis and risk assessment, and the final executable optimized path and corresponding decision support information are output.

[0022] S1: In this step, multidimensional heterogeneous data is first acquired. This data includes different types and dimensions of data such as basic parameters, state indicators, matching scores, and complexity quantification values. Then, it is processed through a series of data preprocessing techniques, such as missing value imputation, outlier detection, and standardization. Next, appropriate encoding or extraction methods are applied to different types of features, such as one-hot encoding of categorical features, statistical extraction of temporal features, and vectorization representation of text features. Finally, a preliminary transformed feature matrix is ​​obtained, laying the foundation for subsequent manifold representation and dimensionality reduction processing.

[0023] S2: In this step, based on the feature matrix of the initial transformation, the Hamiltonian-Poisson integrator technique is introduced. By solving the partial differential equation system, the topological structure of the data is preserved, thereby realizing the isomorphic representation of heterogeneous data and obtaining a unified manifold embedding representation. Then, the geodesic distance matrix between data points in the manifold embedding representation is calculated. The manifold learning algorithm is applied to perform dimensionality reduction based on the geodesic distance matrix, reducing the feature space dimension to the preset target dimension. Finally, a standardized feature vector set is output, providing a high-quality feature representation for subsequent path template matching.

[0024] S3: In this step, based on the standardized feature vector set, a pre-built knowledge graph and standard path template library are extracted from the knowledge base to construct the mapping relationship between features and templates. Then, a stochastic neural network architecture containing randomization layers, multiple hidden layers, attention layers, and output layers is designed and initialized. The network is trained using an improved stochastic gradient descent algorithm. Random field theory is introduced to handle the complex correlations and uncertainties between features. Finally, the matching probability between feature vectors and each path template is calculated through forward propagation and Monte Carlo sampling, and the optimal initial path template and its key node sequence are selected.

[0025] S4: In this step, based on the initial path template and key node sequence, combined with user individual characteristics and system resource constraints, a constrained optimization problem is constructed. The path optimization problem is transformed into an optimal transmission problem in a high-dimensional space. The source distribution representing the initial path and the target distribution representing the ideal path are constructed. The transmission cost matrix between any two points in the state space is calculated. The convex relaxation method is applied to transform the optimal transmission problem into a Kantorovich dual form for solution. The original transmission plan is reconstructed from the dual solution, and the node parameters are finely adjusted to finally obtain a personalized path scheme that meets the constraints.

[0026] S5: In this step, based on the personalized path scheme, real-time monitoring data and feedback information of the system are obtained, the initial state space corresponding to the state and path is constructed, the conditional probability relationship between each path node is modeled using a Bayesian network, the execution probability and priority of each node are calculated, a reinforcement learning algorithm is introduced to learn the optimal path adjustment strategy, a dynamic adjustment decision model is constructed, and then the insertion, deletion, replacement and time adjustment operations of path nodes are automatically executed based on real-time data, and the dynamically updated personalized path is output, so that the path has adaptive capability.

[0027] S6: In this step, based on the dynamically updated personalized path, Monte Carlo simulation parameters and evaluation indicators are set. The Monte Carlo method is applied to generate multiple sets of random disturbance and fluctuation scenarios to simulate the execution of the path under different conditions. The path execution results and performance indicators under each scenario are collected, statistical analysis and risk assessment are performed, potential risk points are identified, and the path is finally adjusted. The final executable optimized path and corresponding decision support information are output to ensure the stability and reliability of the path.

[0028] It is important to note that this method is not a disease diagnosis protocol, but rather a general approach to dynamically generating digital pathways. It does not directly diagnose diseases or make medical decisions, but rather optimizes process execution paths and resource allocation based on existing diagnostic and treatment decisions. In addition to hospital settings, this method is also widely applicable to the following non-medical scenarios:

[0029] 1. Manufacturing Process Optimization: In the field of precision manufacturing, this method can be applied to personalized production path planning for complex products. The system collects multi-dimensional heterogeneous data such as product specifications, raw material characteristics, and production equipment status, constructs standardized feature vectors, matches the most suitable production process template, and optimizes based on current order characteristics and factory resource conditions to generate personalized production paths. When equipment failures or material changes occur during production, the system can dynamically adjust the production path to ensure production efficiency and product quality.

[0030] 2. Logistics Delivery Route Planning: In the logistics industry, this method can be used for intelligent delivery route generation. Based on heterogeneous information such as package characteristics (volume, weight, timeliness requirements), delivery area features, and real-time traffic data, the system matches suitable delivery route templates and optimizes them in conjunction with current vehicle resources and road conditions to generate personalized delivery plans. When encountering road congestion or weather changes, the system can adjust delivery routes in real time, improving delivery efficiency and on-time delivery rates.

[0031] In summary, this method is a general-purpose dynamic digital path generation technology applicable to various scenarios requiring refined process management and resource optimization. It is not only used in hospital patient management but also widely applicable to non-medical fields such as manufacturing, logistics, project management, and education. Its core value lies in achieving personalized path generation, dynamic adjustment, and risk prediction through advanced data processing and optimization algorithms, thereby improving system operational efficiency and service accuracy.

[0032] like Figure 2 As shown in the figure, this embodiment provides a method for dynamically generating digital paths, in which the preprocessing and feature extraction of the multidimensional heterogeneous data to obtain a preliminary transformed feature matrix includes:

[0033] S1.1 receives multidimensional heterogeneous data, which includes basic parameters, state indicators, matching degree scores and complexity quantification values. Through missing value imputation, outlier detection and standardization, it outputs a preprocessed multi-source heterogeneous dataset.

[0034] S1.2 Based on the preprocessed multi-source heterogeneous dataset, perform one-hot encoding on categorical features, extract statistics from temporal features, and vectorize text features to output the feature matrix of the preliminary transformation.

[0035] S1.1: In this step, the system first receives multidimensional heterogeneous data input. These data come from diverse sources and are of various types, mainly including four key data categories: basic parameters (such as numerical and categorical data such as age and gender), which are usually static attributes describing the basic characteristics of the object; state indicators (such as time-series data of various physiological indicators), which usually represent the dynamic characteristics of the object's state changing over time; matching score (such as supply and demand matching similarity data), which usually reflects the degree of fit between the object and the target or resource; and complexity quantification value (such as operation difficulty score), which is usually used to measure the complexity and challenge of a task or process.

[0036] S1.2: After receiving multidimensional heterogeneous data, the system performs comprehensive preprocessing on this data. First, missing value imputation is performed, using different strategies such as mean / median / mode imputation, K-nearest neighbor imputation, or model prediction imputation based on the data type and distribution characteristics. Then, outlier detection is performed, identifying and processing outliers through statistical methods (such as Z-score method, IQR method) or machine learning methods (such as clustering or isolated forest). Finally, standardization is performed to transform data of different dimensions to the same scale, usually using Z-score standardization, Min-Max scaling, or robust scaling. After these processes, the preprocessed multi-source heterogeneous dataset is output, preparing it for subsequent feature extraction.

[0037] This embodiment provides a method for dynamically generating digital paths, in which the standardized feature vector set is obtained based on the feature matrix of the initial transformation through nonlinear manifold representation construction and dimensionality reduction processing, including:

[0038] S2.1 Based on the feature matrix of the preliminary transformation, the Hamiltonian-Poisson integrator technique is introduced to solve partial differential equations to maintain the data topology and output a unified manifold embedding representation;

[0039] S2.2 Based on the unified manifold embedding representation, calculate the geodesic distance matrix between data points in the unified manifold embedding representation, apply the manifold learning algorithm to perform dimensionality reduction processing based on the geodesic distance matrix to reduce the feature space dimension to a preset target dimension, and output the standardized feature vector set.

[0040] S2.1: In this step, based on the characteristic matrix of the initial transformation, the system innovatively introduces the Hamiltonian-Poisson integrator technique to construct a nonlinear manifold representation. This is a data representation method based on differential equations. By solving a specific system of partial differential equations (specifically the Hamiltonian-Poisson equations), it can effectively preserve the topological structure and intrinsic geometric properties of the data during dimensionality reduction and transformation. The system first constructs the local connectivity relationships between data points, forming a discrete approximation of the data manifold; then, based on this, it defines appropriate energy functionals and boundary conditions; next, it solves the corresponding partial differential equations using numerical methods (such as the finite element method or spectral method) to obtain the coordinate mapping of the data points on the low-dimensional manifold; finally, it performs post-processing and normalization on the calculation results, outputting a unified manifold embedding representation. This representation can simultaneously handle different types of heterogeneous data and preserve their intrinsic relational structures.

[0041] S2.2: After obtaining a unified manifold embedding representation, the system calculates the geodesic distance matrix between data points on the embedding manifold. Geodesic distance refers to the shortest path length between two points on the data manifold, and it reflects the true similarity of data on the manifold better than Euclidean distance. When calculating geodesic distance, local connectivity relationships are first constructed based on a K-nearest neighbor graph or an ε-neighborhood graph; then, a shortest path algorithm (such as Dijkstra's algorithm or Floyd-Warshall's algorithm) is applied to calculate the geodesic distance between any two points on the manifold; finally, the geodesic distances between all point pairs are organized into a distance matrix. With the geodesic distance matrix, the system applies manifold learning algorithms (such as Isomap, t-SNE, or UMAP) for dimensionality reduction, reducing the high-dimensional manifold embedding representation to a preset target dimension (usually 10-50 dimensions) while preserving the distance relationships between data points to the greatest extent possible. Finally, the system outputs a standardized feature vector set, which has a unified dimension, scale, and representation, providing high-quality feature input for subsequent path template matching.

[0042] This embodiment provides a method for dynamically generating digital paths. The method involves constructing a mapping relationship between features and path templates using a random neural network based on the standardized feature vector set and a pre-built knowledge graph and standard path template library. The model is then trained to output an initial path template and the corresponding key node sequence. This includes:

[0043] S3.1 Receive the standardized feature vector set, extract the pre-built knowledge graph and standard path template library from the knowledge base, establish the correspondence between the standardized feature vector set and the path template through association mapping, and output the training dataset;

[0044] S3.2 Based on the training dataset, construct a randomized neural network architecture including a randomization layer, a hidden layer, an attention layer, and an output layer; set the network parameters and learning rate and perform random initialization of the weights; output the initialized randomized neural network model.

[0045] S3.3 Based on the initialized stochastic neural network model and the training dataset, the stochastic gradient descent algorithm is applied to train the model. Random field theory is introduced to construct a Markov random field model to represent the conditional independence structure between features, and the trained stochastic neural network model is output.

[0046] S3.4 Based on the trained stochastic neural network model, combined with the standardized feature vector set, the matching probability between the standardized feature vector set and each standard path template in the standard path template library is calculated through forward propagation and Monte Carlo sampling. The path template with the highest matching probability value is selected, and the initial path template and the corresponding key node sequence are output.

[0047] S3.1: In this step, the system receives a standardized feature vector set and extracts two types of key information from a pre-built knowledge base: a domain knowledge graph and a standard path template library. The domain knowledge graph is a structured form of knowledge representation that constructs a conceptual network for a specific domain through entities (nodes) and relationships (edges). It includes multiple core entities such as parameter feature entities, path node entities, and resource entities, as well as various relationships between them such as "influence," "dependency," and "need." The standard path template library contains a predefined set of path templates. Each template consists of a series of ordered path nodes with specific execution order and time constraints, along with applicable conditions and historical performance records. The system maps the standardized feature vector set to the path template library through the relationships in the knowledge graph, constructing a training triplet dataset in the form of (feature vector, path template, matching score), providing a foundation for subsequent neural network training.

[0048] S3.2: Based on the training dataset output from the previous step, the system constructs a specialized stochastic neural network architecture. Unlike traditional deterministic neural networks, this architecture introduces multi-layered randomness elements, making it more suitable for handling complex feature-template mapping relationships. The network includes: an input layer (number of nodes equal to the feature vector dimension); a special randomization layer (enhancing robustness by adding random noise or applying a random mask); multiple deep hidden layers (typically 3-5 layers, equipped with non-linear activation functions and dropout mechanisms); an attention layer (dynamically focusing on key dimensions in the feature vector); and an output layer (number of nodes equal to the number of templates in the template library). The system sets key parameters such as the learning rate (adaptive strategy, initial value 0.001-0.005), batch size (between 32-128), and weight initialization method (Xavier or He initialization), outputting a structurally complete, parameter-initialized stochastic neural network model.

[0049] S3.3: Based on the initialized stochastic neural network model and training dataset, the system applies an improved stochastic gradient descent algorithm for model training. Specific improvements include: replacing traditional SGD with the Adam optimizer, combined with momentum and adaptive learning rate adjustment; introducing a periodic learning rate adjustment strategy to make the learning rate change periodically within a preset range; and implementing gradient clipping technology to prevent gradient explosion. During optimization, the system uses the Adam optimizer instead of traditional SGD because it can automatically adjust the learning rate for each parameter, accelerating convergence. For example, for parameters with small feature weights, Adam automatically increases their learning rate, while for frequently updated parameters, it decreases the learning rate, making the training process more stable. The system also implements periodic learning rate adjustment, with the learning rate changing periodically between 0.001 and 0.005, helping the model escape local optima. When a gradient value is detected to exceed a preset threshold of 10, the system automatically performs gradient clipping, scaling the gradient value to a reasonable range, effectively preventing gradient explosion during training. Simultaneously, the system innovatively introduces random field theory to handle the complex correlations between input features, constructing a Markov random field model to represent the conditional independence structure between features. Specifically, the system treats input features as random variables and establishes conditional probability dependency graphs between features. For example, in medical scenarios, there are complex conditional dependencies between features such as a patient's age, blood pressure, and heart rate. Through random field modeling, the system can capture complex relationships such as "the conditional correlation between blood pressure and heart rate given an age," enabling the network to consider the mutual influence between features rather than simply treating them as independent inputs. The training process uses a cross-entropy loss function and introduces an L2 regularization term to prevent overfitting, with the regularization coefficient set to 0.001. The system also implements an early stopping strategy, automatically stopping training when the loss on the validation set does not improve for five consecutive epochs to avoid overfitting the training data. Through the comprehensive application of these optimization techniques, the system ultimately outputs a trained random neural network model, which exhibits high accuracy and good generalization ability on feature-template mapping tasks.

[0050] S3.4: Based on a pre-trained stochastic neural network model, the system inputs the standardized feature vector of the current user into the network and calculates the matching probability of each standard path template through forward propagation. To improve decision reliability, the system does not perform a single forward propagation but instead performs multiple Monte Carlo samplings (usually 20-50 times), introducing different random perturbations in the randomization layer each time, and then averaging all sampling results to obtain a more stable matching probability distribution. During template selection, the system adopts a two-stage strategy: first, it retains the top K (usually 3-5) templates with the highest matching probabilities as a candidate set; then, it performs a secondary evaluation on each template in the candidate set, considering factors such as historical success rate, resource compatibility, and complexity, selecting the template with the highest comprehensive score as the initial path template, and extracting its key node sequence, including the functional definition, execution conditions, and expected output of each node.

[0051] This embodiment provides a method for dynamically generating digital paths, in which the following steps are performed: training the model using a stochastic gradient descent algorithm based on the initialized stochastic neural network model and the training dataset; introducing random field theory to construct a Markov random field model representing the conditional independence structure between features; and outputting the trained stochastic neural network model. The method includes:

[0052] S4.1 Based on the initialized stochastic neural network model and the training dataset, the model is trained using the Adam optimizer combined with a periodic learning rate adjustment strategy and gradient pruning technique to obtain the basic training model;

[0053] S4.2 Based on the basic training model, construct a Markov random field model to represent the conditional independence structure between features, establish a conditional probability dependency graph between features and learn the mutual influence between features, and output the trained random neural network model.

[0054] S4.1: In this step, the system trains the model using the advanced Adam optimizer based on the initialized stochastic neural network model and training dataset. The Adam optimizer combines the advantages of momentum gradient descent and RMSProp, automatically adjusting the learning rate for each parameter, making the training process more efficient. The system further introduces a cyclic learning rate strategy, allowing the learning rate to periodically change between a predetermined minimum and maximum value. This strategy helps the model escape local optima and converge to the global optimum more quickly. Simultaneously, the system implements gradient pruning, cutting off gradients when they exceed a preset threshold, effectively preventing gradient explosion and improving training stability. Through the combined application of these advanced techniques, the system can efficiently train complex stochastic neural network models, ultimately obtaining a basic training model, laying the foundation for the next step of introducing random field theory.

[0055] S4.2: Based on the basic training model obtained in the previous step, the system innovatively introduces a Markov random field model to represent the conditional independence structure between features. Specifically, the system treats input features as random variable nodes, and then establishes a conditional probability dependency graph based on the correlation and causal relationships between features. This graph precisely describes which features directly affect other features and which features are independent of each other given certain features. The system automatically constructs this dependency graph structure and learns the conditional probability distribution between features by analyzing training data and model learning results. Subsequently, the system integrates this Markov random field model with a neural network model, enabling the network to consider the complex interactions between features when making predictions, rather than simply treating features as independent inputs. This integration allows the model to more accurately capture complex patterns and relationships in the high-dimensional feature space, improving the model's predictive power and robustness. Finally, the system outputs a trained stochastic neural network model enhanced with random field theory, which can more effectively handle complex correlations and uncertainties between features.

[0056] In one embodiment, during the training of the stochastic neural network model, the system employs an advanced combination of optimization strategies, fully considering the conditional independence structure between features, thereby achieving accurate modeling of complex feature relationships.

[0057] First, the system trains the model using the Adam optimizer based on an initialized stochastic neural network structure and a pre-prepared training dataset. The Adam optimizer, a widely used optimization algorithm in deep learning, combines the advantages of momentum gradient descent and RMSProp, adaptively calculating different learning rates for each parameter. This adaptability makes parameter updates more accurate, making it particularly suitable for handling complex situations such as sparse gradients and non-stationary objectives. In the actual implementation, the system sets the initial learning rate to 0.001 and uses the classic parameter configuration of β1=0.9 and β2=0.999, maintaining training stability while ensuring convergence speed.

[0058] The system further introduces a periodic learning rate adjustment strategy, which allows the learning rate to change periodically between a predefined minimum value (usually 1 / 10 of the base learning rate) and a maximum value (usually 10 times the base learning rate) according to a triangular or cosine function pattern.

[0059] Simultaneously, the system implements gradient pruning to prevent gradient explosion. Specifically, when the calculated gradient norm exceeds a preset threshold (typically set to 5.0), the system proportionally reduces the norm of each component of the gradient vector until it equals the threshold. This technique is particularly important when processing recurrent neural network components and long sequence data, effectively improving the stability of the training process. For example, in a hospital patient treatment pathway planning application, when the system processes patient data containing long-term medical histories, gradient pruning reduced the frequency of gradient explosion events during training by approximately 90%, significantly improving the model training success rate.

[0060] After obtaining the basic training model, the system innovatively introduces a Markov random field model to enhance its ability to model complex relationships between features. A Markov random field is an undirected graphical model that can effectively represent the conditional independence structure between random variables. In the system implementation, each input feature is treated as a node in the random field, and the edges between nodes represent the direct correlation between features. By analyzing the conditional mutual information of features in the training data, the system automatically identifies significant conditional dependencies and constructs a conditional probability dependency graph.

[0061] For example, in manufacturing production path optimization applications, the system might discover a direct correlation between raw material purity and processing temperature. However, given these two factors, the relationship between production equipment model and product yield might exhibit conditional independence. The system automatically captures and represents this complex conditional independence structure, avoiding the model learning spurious correlations. Through conditional probability dependency graphs, the system can learn a more accurate joint probability distribution, improving the effectiveness of feature representation.

[0062] The system deeply integrates a constructed Markov random field model with a neural network model. Specifically, a random field structure is introduced into the intermediate layers of the neural network, allowing the feature representations to consider both the independent contribution of individual features and the conditional dependencies between features. This integrated architecture employs an end-to-end training approach, simultaneously optimizing the parameters of the neural network and the structural parameters of the random field, resulting in mutual enhancement. In logistics delivery route planning applications, this integrated method improves route prediction accuracy by more than 15%, particularly when handling highly correlated feature groups (such as weather conditions, traffic conditions, and delivery times).

[0063] Ultimately, the system outputs a trained stochastic neural network model enhanced with random field theory. This model not only handles complex patterns in high-dimensional feature spaces but also accurately captures the conditional independence and dependencies between features, significantly improving prediction performance and generalization ability. In complex project management applications, this enhanced model reduced the average error in project schedule prediction from 18% of the traditional model to 7%, demonstrating a clear advantage, particularly when dealing with multiple interdependent project factors.

[0064] Through this advanced model training and enhancement strategy, the system can fully explore the complex relationships between features, providing a strong mathematical foundation for subsequent path template matching and dynamic adjustment, thereby achieving accurate processing of high-dimensional heterogeneous data and intelligent generation of complex paths.

[0065] This embodiment provides a method for dynamically generating digital paths. Based on the trained stochastic neural network model and combined with the standardized feature vector set, the matching probability between the standardized feature vector set and each standard path template in the standard path template library is calculated through forward propagation and Monte Carlo sampling. The path template with the highest matching probability is selected, and the initial path template and its corresponding key node sequence are output, including:

[0066] S5.1 Based on the trained stochastic neural network model and the standardized feature vector set, the number of Monte Carlo samplings is set as a preset threshold, the forward propagation calculation of the preset threshold number of times is performed, and the average of all sampling results is taken to obtain the stable matching probability distribution between the standardized feature vector set and each standard path template in the standard path template library.

[0067] S5.2 Based on the stable matching probability distribution, the standard path templates in the standard path template library are sorted from high to low according to the matching probability value. The top K standard path templates with the highest matching probability are retained as the candidate standard path template set. Each standard path template in the candidate standard path template set is subjected to secondary evaluation, and the standard path template corresponding to the maximum comprehensive score is selected. The initial path template and the corresponding key node sequence are output.

[0068] S5.1: In this step, the system uses a Monte Carlo sampling method for path template matching based on a trained stochastic neural network model and the user's standardized feature vector set. The system first sets the number of Monte Carlo samplings to a preset threshold (usually between 20-50 times), which needs to balance computational efficiency and result stability. For each sampling, the system inputs the feature vector into the stochastic neural network, introduces different random perturbations (such as different Gaussian noise or random masks) into the randomization layer, and then performs a complete forward propagation calculation to obtain the matching probability of each standard path template under that sampling. This process is repeated until the preset number of samplings is reached. Then, the system takes the arithmetic mean of all sampling results to obtain the final stable matching probability distribution. This multiple sampling strategy effectively reduces the randomness and uncertainty of a single prediction, providing more stable and reliable template matching results.

[0069] S5.2: Based on the stable matching probability distribution obtained in the previous step, the system first sorts all templates in the standard path template library from high to low according to their matching probability values. Then, the system retains the top K templates (K is usually 3-5) as the candidate standard path template set. This soft selection mechanism avoids the risks that may arise from relying solely on a single template with the highest probability. Next, the system performs a secondary evaluation on each template in the candidate set, considering multiple factors, including: the template's historical success rate (based on past application records), the template's compatibility with current resource constraints (checking whether resource requirements can be met), and the template's complexity and operability (assessing implementation difficulty and risk). The system calculates a comprehensive score for each candidate template based on these factors and selects the template with the highest score as the final initial path template. The system then extracts the key node sequence from this template, including the specific functional definition, execution conditions, expected output, time window, and resource requirements of each node, as the basis for subsequent personalized optimization. Through this two-stage selection strategy, the system can select the path template most suitable for the current user characteristics and system environment while ensuring reliability.

[0070] During the path template matching stage, the system uses Monte Carlo sampling techniques and a multi-level evaluation strategy to ensure that the most suitable path template for the current feature vector is found in the high-dimensional feature space, while taking into account the uncertainty of model prediction and various factors of actual application scenarios.

[0071] The system first uses a trained stochastic neural network model, combined with the standardized feature vector set to be processed, to set a preset threshold for the number of Monte Carlo samplings, typically between 20 and 50. The selection of this threshold requires balancing computational resource consumption and prediction stability. In practical applications, the system dynamically adjusts this parameter based on the importance of the task and available computational resources. For example, in high-risk medical route planning scenarios, the number of samplings might be set to 50 to ensure high reliability of the results; while in logistics and delivery planning with high real-time requirements, it might be set to 20 to ensure response speed.

[0072] For each Monte Carlo sampling, the system introduces different random perturbations into the randomization layer of the stochastic neural network. These perturbations can be the addition of different Gaussian noises (typically with a mean of 0 and a standard deviation of 0.1 to 0.3) or the application of different random dropout masks (typically with a dropout rate of 0.1 to 0.5). This introduction of random perturbations simulates small changes in the feature data and the uncertainty of the model itself, helping to improve the robustness of the prediction results. For example, in manufacturing production path planning, even with small fluctuations in raw material parameters, the system can provide stable path recommendations.

[0073] The system performs a complete forward propagation calculation on the feature vector for each randomly perturbed version, obtaining the matching probability between the feature vector and each template in the standard path template library under the current sampling. This probability distribution reflects the fit between each path template and the input feature under the current random state. Subsequently, the system performs an arithmetic mean of all the probability distributions obtained from sampling to obtain the final stable matching probability distribution. This method of averaging multiple samplings significantly reduces the impact of random fluctuations in a single prediction, providing more reliable template matching results. In practical applications of complex project management, this method improves prediction accuracy by approximately 18% compared to single predictions, especially for feature vectors at template boundary conditions.

[0074] Based on the obtained stable matching probability distribution, the system sorts all templates in the standard path template library from high to low according to their matching probability values. This sorting process fully utilizes the feature-template mapping relationship captured by the stochastic neural network model, reflecting the adaptability of each template at the pure feature matching level. Then, the system retains the top K templates with the highest ranking as a candidate standard path template set, where K is typically set to 3 to 5. For example, in personalized learning path planning for education and training, the system may retain the 5 templates with the highest matching probability from more than 100 standard course path templates as a candidate set. This "soft selection" mechanism avoids the risks that may arise from relying solely on a single template with the highest probability, increases system fault tolerance, and maintains computational efficiency.

[0075] For the selected set of candidate standard path templates, the system conducts a more comprehensive and in-depth secondary evaluation. This evaluation not only considers the matching probability predicted by the model but also incorporates various practical application factors. First, there is a historical success rate evaluation, where the system analyzes the proportion of successful cases and performance of each candidate template in historical applications. For example, in hospital patient pathway planning, the system considers the historical cure rate and complication rate of each treatment pathway template in similar patient groups. Second, there is a resource compatibility evaluation, where the system checks whether the available resources in the current environment can meet the needs of each candidate template. In logistics and distribution scenarios, this may involve a matching analysis of currently available vehicle types, warehouse capacity, and human resources. Third, there is a template complexity and operability evaluation, where the system assesses the difficulty and risks of implementing each candidate template. For example, in manufacturing production pathway planning, the system considers process complexity, technical proficiency requirements, and potential production risks.

[0076] The system calculates a comprehensive score for each candidate template based on these evaluation dimensions. This score is typically calculated using a weighted summation method, with different weights assigned to each evaluation dimension depending on the application scenario. For example, in high-risk medical scenarios, historical success rate may have a larger weight; while in resource-constrained production environments, resource compatibility may have a higher weight. The system ultimately selects the template with the highest comprehensive score as the initial path template. This selection considers both model predictions and practical application factors, ensuring an organic combination of theory and practice.

[0077] After selecting an initial path template, the system extracts the sequence of key nodes from that template, including core information such as the specific functional definition, execution conditions, expected output, time window, and resource requirements for each node. This detailed node information provides the foundational data for subsequent personalized optimization and dynamic adjustment of the path. For example, in complex project management applications, the system extracts information such as milestone nodes, delivery requirements, resource allocation plans, and key risk points for each stage of the project to form a complete initial project path.

[0078] By employing this path template matching strategy based on Monte Carlo sampling and multi-level evaluation, the system can consider both the statistical stability of model predictions and various factors in practical applications when processing high-dimensional heterogeneous data. This provides a high-quality initial template for the next step of personalized path optimization, thereby achieving intelligent generation of digital paths.

[0079] This embodiment provides a method for dynamically generating digital paths. The method involves transforming the path optimization problem into a source-to-target distribution transmission problem using optimal transmission theory, based on the initial path template and corresponding key node sequence, combined with user individual characteristics and resource constraints. This transformation is then solved and optimized to obtain a personalized path solution.

[0080] S6.1 Receive the initial path template and the corresponding key node sequence, obtain user individual characteristics and system resource constraints, construct an optimization problem including objective function, decision variables and constraints, and output the constrained optimization problem definition;

[0081] S6.2 Based on the definition of the constrained optimization problem, construct the source distribution representing the initial path template and the target distribution representing the ideal personalized path, calculate the transmission cost matrix between any two points in the state space and initialize the transmission plan, and output the initial solution of the transmission plan;

[0082] S6.3 Based on the initial solution of the transmission plan, the optimal transmission problem is transformed into a Kantorovich dual form by applying the convex relaxation method. The dual problem is solved iteratively and the original transmission plan is reconstructed from the dual solution to output the optimized path scheme.

[0083] S6.4 Based on the optimized path scheme, a sensitivity analysis model for node parameter adjustment is constructed to optimize the execution time, resource allocation, and dependencies of nodes in the optimized path scheme, and the personalized path scheme is output.

[0084] S6.1: In this step, the system receives the initial path template and its key node sequence, and simultaneously acquires user individual characteristics and system resource constraints to construct a mathematically rigorous definition of the optimization problem. The system first represents the initial path template as a directed acyclic graph (DAG), where nodes represent operation steps in the path, edges represent dependencies between nodes, and each node has multiple attributes such as execution time, resource requirements, and priority. Simultaneously, the system formalizes various constraints: user individual characteristic constraints (such as time sensitivity, resource preferences, etc., usually represented as soft constraints); system resource constraints (the quantity and time distribution of available resources, represented as hard constraints); time window constraints (some operations must be completed within a specific time window); and node dependency constraints (defining the order in which nodes are executed). Based on these elements, the system constructs a multi-objective optimization problem. The objective function includes multiple components such as time efficiency, resource utilization, expected results, and individual fit, which are combined into a comprehensive objective function through weighted coefficients. Finally, the system outputs a complete definition of the constrained optimization problem, providing a foundation for subsequent optimization solutions.

[0085] S6.2: Based on the optimization problem definition in the previous step, the system innovatively transforms the path optimization problem into an optimal transport problem in a high-dimensional space. Within this framework, the system defines a source distribution μ (representing the state distribution of the initial path template) and a target distribution ν (representing the ideal personalized path state distribution), as well as a transport cost function c(x,y) (representing the "cost" required to move from state x to state y). In the specific implementation, the system first constructs a discretized state space; then calculates the probability quality distribution μ of the initial path template in the state space; next, based on individual user characteristics and the system optimization objective, it constructs the probability quality distribution ν of the ideal path; subsequently, it calculates the transport cost matrix C between any two points in the state space; finally, it initializes the transport plan π as a joint distribution satisfying marginal constraints (such as the multiplicative distribution of μ and ν). To efficiently solve large-scale optimized transport problems, the system employs approximation algorithms such as regularization, the Sinkhorn algorithm, and multi-resolution strategies, ultimately outputting the initial solution of the transport plan, laying the foundation for the next optimization step.

[0086] S6.3: Based on the initial solution of the transmission plan, the system applies a convex relaxation method to transform the optimal transmission problem into a Kantorovich dual form for solution. The basic idea of ​​this method is to transform the original non-convex optimization problem into a more easily solvable convex optimization problem through relaxation techniques. Specifically, the system first expresses the original optimal transmission problem as minimizing the sum of the products of the transmission cost and the transmission plan, with the constraint that the marginal distribution of the transmission plan equals the source and target distributions; then it transforms it into the Kantorovich dual form, expressing it as maximizing the sum of the products of the dual variables and their distributions, with the constraint that the sum of the dual variables is less than or equal to the transmission cost. The system solves the dual problem through an iterative computation process: first, the dual variables are initialized as zero vectors; then, one dual variable is fixed while the other is solved, iterating repeatedly until convergence; finally, the original transmission plan is reconstructed from the dual solution. To handle high-dimensional, large-scale problems, the system employs acceleration techniques such as problem decomposition, parallel computing, sparse matrix optimization, and adaptive precision control, ultimately outputting the optimized path scheme.

[0087] S6.4: Based on the optimized path scheme, the system performs fine-tuning at the node level, further adjusting the specific parameters of each node. The system focuses on optimization in four dimensions: time dimension (adjusting the start time, duration, and completion deadline of nodes); resource dimension (optimizing resource type selection and quantity allocation); dependency dimension (fine-tuning the dependencies and parallel strategies between nodes); and content dimension (adjusting the specific execution content of nodes based on user characteristics). The system first constructs a sensitivity analysis model for node parameter adjustments. By slightly perturbing each parameter and calculating the path performance changes, a sensitivity coefficient is obtained. Then, parameters with high sensitivity are adjusted first. The system pays special attention to nodes on critical paths, resource-intensive nodes, and high-risk nodes, continuously verifying whether the modified path still meets all constraints during the optimization process. Finally, the system outputs a personalized path scheme that meets the constraints, including a complete set of path nodes (each node has precisely defined execution parameters), a dependency graph between nodes, a detailed resource allocation plan, and key performance indicator (KPI) predictions.

[0088] Specifically, personalized route optimization is the core of this method. The system mathematically transforms the route optimization problem using optimal transport theory, achieving a precise conversion from standard templates to personalized solutions. This process not only considers the unique needs of users but also takes into account system resource constraints, ultimately generating highly customized and practically feasible route solutions.

[0089] In the first step, the system receives the initial path template and its key node sequence output from the previous stage, and simultaneously acquires the current user's individual characteristics and the system's resource constraints. User individual characteristics include feature data directly related to path execution; for example, in a medical scenario, this might include a patient's complication history, drug allergy history, and preferred treatment periods; in a manufacturing scenario, it might include specific material requirements for the product, customer-specified delivery dates, and quality standards. System resource constraints describe the objective limitations of the path execution environment, such as the number of hospital beds, specialist physician scheduling, and available time periods for special equipment; or the production equipment capacity, technical personnel allocation, and raw material inventory status of a factory.

[0090] Based on these inputs, the system constructs a formalized optimization problem. First, the objective function is defined, typically a combination of multiple objectives, including dimensions such as path execution efficiency (e.g., minimum total completion time), resource utilization (e.g., load balancing of critical resources), user experience (e.g., minimum waiting time), and result quality (e.g., maximum expected success rate). Each dimension is assigned different weights based on the application scenario. For example, in emergency medical scenarios, time efficiency may have a higher weight; while in precision manufacturing scenarios, quality indicators may be more important. Decision variables mainly include adjustable parameters such as the execution time of each node in the path, the execution order, and the type and quantity of allocated resources. Constraints include hard constraints (conditions that must be met, such as drug safety interval requirements in medical scenarios or process temperature ranges in manufacturing scenarios) and soft constraints (conditions that should be met as much as possible but can be appropriately relaxed, such as patient preferences for specific examination time periods or completion time requirements for non-critical procedures).

[0091] After defining the problem, the system innovatively applies Optimal Transport Theory for solution. Originally proposed by the French mathematician Monge and later extended by Kantorovich, this theory is a mathematical framework for studying how to transform one probability distribution into another at minimal cost. In this method, the system represents the initial path template as the source distribution, with each node possessing a specific probability quality reflecting its importance in the standard path; the ideal personalized path is represented as the target distribution, whose probability quality distribution reflects the ideal state considering user characteristics and resource constraints.

[0092] The system then constructs a transmission cost matrix between any two points in the state space. This matrix quantifies the cost of transforming a node in the source distribution into a node in the target distribution. For example, in a medical scenario, changing the execution time of a specific examination may result in additional waiting time or resource rescheduling costs; in a production scenario, adjusting the sequence of processes may lead to equipment reconfiguration costs or product quality risks. These costs are quantified through comprehensive analysis of historical data and expert knowledge, forming a complete cost matrix. The system initializes a feasible transmission plan, typically employing a nearest neighbor strategy or a proportional allocation strategy, providing a starting point for iterative optimization.

[0093] In the third step, the system applies convex relaxation to transform the optimal transport problem into a Kantorovich dual form. This mathematical transformation converts the originally complex nonlinear optimization problem into a more easily solvable linear programming problem. The core idea of ​​the Kantorovich dual form is to introduce potential functions and indirectly solve the optimal transport plan by maximizing the expected difference of these potential functions. The system employs alternating iterative numerical algorithms to solve the dual problem, such as the Sinkhorn algorithm, which significantly improves computational efficiency by adding an entropy regularization term, making it possible to solve large-scale practical problems. In practical applications, the system typically sets appropriate termination conditions, such as the relative change of solutions between two consecutive iterations being less than a preset threshold (e.g., 0.001) or reaching the maximum number of iterations (e.g., 200 iterations).

[0094] After the algorithm converges, the system reconstructs the original transport plan from the dual solution, obtaining the optimal transformation scheme from the standard path template to the personalized path. This scheme specifies how to adjust the nodes in the initial path, including which nodes need to be retained, deleted, replaced, or reordered, and how to adjust the specific parameters of each node. For example, in a patient treatment pathway, the system may suggest advancing a certain examination, postponing a certain drug treatment, or adding additional monitoring nodes based on the patient's specific condition; in a production pathway, the system may suggest adjusting the processing parameters of a certain step, or replacing specific equipment and processes for special materials.

[0095] In the final step, the system further refines and performs sensitivity analysis on the optimized path scheme. The sensitivity analysis model assesses the impact of changes in parameters at each node in the path on the overall objective function, identifying key parameters and sensitive points. For example, the system might find that in a medical path, the timing of a specific examination significantly affects the overall treatment outcome; or in a manufacturing path, temperature control of a certain heat treatment process is particularly critical to the final product quality. Based on the sensitivity analysis results, the system fine-tunes the execution time, resource allocation, and dependencies of nodes in the path scheme.

[0096] In terms of execution time, the system considers the flexibility of time windows and the precision of time points. For example, some medical examinations must be performed at fixed times, while some rehabilitation activities can be scheduled within a wider time window. In terms of resource allocation, the system optimizes the resource type, quantity, and usage duration of each node to ensure that critical resources are prioritized while avoiding resource idleness. In terms of dependency, the system adjusts the logical relationships between nodes to ensure that necessary sequential constraints are met, while maximizing the possibility of parallel execution to improve overall efficiency.

[0097] Through this series of meticulous optimization steps, the system ultimately outputs a complete personalized path solution, including detailed node execution plans, resource allocation tables, and key control parameters. This solution is highly adaptable to individual user characteristics while fully considering the actual constraints of system resources, achieving an organic unity between theoretical optimization and practical feasibility. In hospital patient management scenarios, this optimization may shorten the average treatment cycle for patients by 15-20% while improving resource utilization by 10-15%; in manufacturing scenarios, it may shorten product delivery cycles by 10-15% while reducing resource consumption by 5-10%. Regardless of the application scenario, this personalized path optimization method based on optimal transport theory demonstrates significant efficiency improvements and resource-saving potential.

[0098] This embodiment provides a method for dynamically generating digital paths, in which the optimal transmission problem is transformed into a Kantorovich dual form by applying a convex relaxation method based on the initial solution of the transmission plan, and the original transmission plan is reconstructed from the dual solution through iterative solution to output an optimized path scheme, including:

[0099] S7.1 Based on the initial solution of the transmission plan, initialize the Kantorovich dual variables, and use problem decomposition and parallel computing techniques to iteratively solve the problem and obtain the final dual solution;

[0100] S7.2 Based on the final dual solution, reconstruct the original transmission plan and generate the modified path node parameters, and output the optimized path scheme.

[0101] S7.1: In this step, the system begins solving the Kantorovich dual problem based on the initial solution of the transport plan. First, the system initializes a pair of dual variables, α and β, as zero vectors (or other suitable initial values), corresponding to each state point in the source and target distributions, respectively. Then, the system employs a problem decomposition technique to break down the entire large-scale optimization problem into a series of smaller subproblems, each dealing with a local region of the state space. This decomposition strategy significantly reduces computational complexity, making the large-scale problem tractable. Next, the system solves these subproblems in parallel on a multi-core processor or in a distributed computing environment, further improving computational efficiency. In each iteration step, the system alternately fixes one set of dual variables while optimizing another set, for example, fixing β to optimize α, then fixing α to optimize β. This process continues until convergence conditions are met: the change in the dual objective function is less than a preset threshold, the violation of the original constraints is less than a preset threshold, or the maximum number of iterations is reached. Through this efficient iterative solution process, the system finally obtains a stable dual solution, preparing for the next step of reconstructing the original transport plan.

[0102] S7.2: Based on the final dual solution obtained in the previous step, the system begins to reconstruct the original optimal transmission plan. Specifically, the system uses the dual complementary relaxation condition to identify state pairs (i,j) that satisfy α(i) + β(j) = c(i,j). These state pairs are precisely the state pairs that should have positive traffic in the optimal transmission plan. The system then constructs the original transmission plan accordingly. Ensure that it satisfies all marginal constraints (i.e. The sum of rows equals the source distribution μ, and the sum of columns equals the target distribution ν. With the optimal transmission plan, the system further extracts path transformation information to determine which initial path states should be transformed into which target path states, and the degree of transformation. Based on this information, the system generates modified path node parameters, including the optimized execution time of each node, resource allocation scheme, and inter-node dependencies. These parameters comprehensively consider individual user characteristics and system resource constraints, realizing the transformation from the initial path template to a personalized optimized path. Finally, the system outputs an optimized path scheme that, while maintaining the core functionality of the initial path template, achieves personalized adaptation to user needs through rigorous mathematical optimization based on optimal transmission theory.

[0103] This embodiment provides a method for dynamically generating digital paths. The method involves, based on the personalized path scheme and combined with real-time monitoring data and feedback information, modeling the conditional probability relationships between nodes using a Bayesian network and constructing a dynamic adjustment decision model using a reinforcement learning algorithm. This model automatically executes path node adjustment operations and outputs a dynamically updated personalized path. The method includes:

[0104] S8.1 Receives the personalized path scheme, obtains real-time monitoring data and feedback information, and outputs the initial state space corresponding to the state-path;

[0105] S8.2 Based on the initial state space, apply a Bayesian network to model the conditional probability relationship between nodes in the personalized path scheme, calculate the execution probability and priority of each path node, and output the node probability distribution model;

[0106] S8.3 Based on the node probability distribution model, a reinforcement learning algorithm is introduced to learn the path adjustment strategy and output a dynamic adjustment decision model;

[0107] S8.4 Based on the dynamic adjustment decision model, combined with the real-time monitoring data and feedback information, automatically execute the insertion, deletion, replacement and time adjustment operations of path nodes in the personalized path scheme, and output the dynamically updated personalized path.

[0108] S8.1: In this step, the system receives the personalized path plan and simultaneously acquires real-time monitoring data and feedback information from the system's operating environment. This real-time data may include current status indicators (such as resource usage, progress completion, quality indicators, etc.), user behavior data, and information on changes in the external environment. The system associates and integrates the personalized path plan with this real-time data to construct an initial state space corresponding to the state and path. This state space is a high-dimensional structure, where each dimension represents a key attribute of the system or environment, each point represents a possible system state, and the path connecting these states represents the possible evolutionary trajectory of the system. This initial state space provides the basic data structure for subsequent Bayesian network modeling and reinforcement learning, enabling the system to make adaptive adjustments in a dynamically changing environment.

[0109] S8.2: Based on the initial state space, the system applies Bayesian network technology to model the conditional probability relationships between nodes in the personalized path scheme. A Bayesian network is a probabilistic graphical model that represents the conditional dependencies between random variables through a directed acyclic graph. In this model, each path node is treated as a random variable, and the edges between nodes represent conditional dependencies. The system learns these conditional probability distributions by analyzing historical data and current real-time information. For example, P(Node B succeeds | Node A succeeds) represents the probability that Node B will succeed given that Node A succeeds. Based on the constructed Bayesian network, the system can calculate the conditional execution probability and priority of each path node in the current state. These probabilities reflect the uncertainty and importance of each node's execution under existing conditions and known information. Finally, the system outputs a node probability distribution model, which accurately captures the uncertainties and risks in the path execution process, providing a probabilistic basis for subsequent path adjustment decisions.

[0110] S8.3: Based on the node probability distribution model, the system introduces a reinforcement learning algorithm to learn the optimal path adjustment strategy. Reinforcement learning is a machine learning method that gradually learns the optimal decision-making strategy through trial and error. The system constructs the path adjustment problem as a Markov Decision Process (MDP): the state space is the set of possible system states (including current execution progress, resource status, etc.); the action space is the set of possible adjustment operations (such as inserting, deleting, replacing nodes, or adjusting time); the reward function reflects the contribution of the adjustment behavior to the goal; and the transition probability is determined by the dynamics of the environment. Based on advanced reinforcement learning algorithms such as Deep Q-Learning (DQN) or Policy Gradient, the system continuously tries different adjustment strategies through interaction with the environment, learning the optimal strategy that maximizes long-term cumulative rewards. During the learning process, the system balances the exploration of new strategies with the utilization of known good strategies, ensuring that it can both discover potentially better strategies and stably execute known effective strategies. Finally, the system outputs a trained dynamic adjustment decision model, which can automatically determine the optimal path adjustment strategy based on the current state.

[0111] S8.4: Based on a dynamically adjusted decision-making model, combined with continuously updated real-time monitoring data and feedback information, the system can automatically execute various adjustment operations in the personalized path plan. These operations include: inserting new nodes (when necessary steps are missing); deleting nodes (when certain steps become unnecessary or redundant); replacing nodes (when a more suitable alternative is found); and time adjustments (optimizing the start and end times of each node). The system selects the optimal adjustment operation based on the current state and learned strategies, and executes these adjustments in real time. After adjustment, the system continues to monitor the execution effect and continuously optimizes the decision-making model and adjustment strategies based on new feedback. This closed-loop feedback mechanism enables the path to be adaptive, able to cope with various changes and challenges encountered during execution. Ultimately, the system outputs a dynamically updated personalized path that not only personally matches user needs but also intelligently adjusts according to actual execution conditions, ensuring efficient achievement of the goal.

[0112] This embodiment provides a method for dynamically generating digital paths. Based on the dynamically updated personalized path, Monte Carlo simulation technology is applied to generate a simulation scenario and run the path simulation. Adjustments are made through statistical analysis and risk assessment, and the final executable optimized path and corresponding decision support information are output, including:

[0113] S9.1 Receives the dynamically updated personalized path, sets the Monte Carlo simulation parameters and evaluation metrics, and outputs the initial simulation configuration;

[0114] Based on the initial simulation configuration, S92 uses the Monte Carlo method to generate multiple sets of random disturbance and fluctuation scenarios, and outputs a simulation scenario set;

[0115] S9.3 Based on the simulation scenario set, perform simulation runs on the dynamically updated personalized paths and collect path execution results and performance indicators under each scenario, and output simulation result dataset;

[0116] S9.4 Based on the simulation result dataset, perform statistical analysis and risk assessment, identify potential risk points, adjust the dynamically updated personalized path, and output the final executable optimized path and corresponding decision support information.

[0117] S9.1: In this step, the system receives dynamically updated personalized paths and sets key parameters and evaluation metrics for the Monte Carlo simulation. Key parameters include: the number of simulation scenarios (typically between 100-1000, depending on problem complexity and computational resources); the magnitude of random perturbations (defining the range of random fluctuations for each variable); the time granularity (defining the smallest time unit for the simulation); and the resource fluctuation model (describing the random variation rules of resource availability). Evaluation metrics include: completion time distribution (reflecting the uncertainty of path execution time); resource utilization (evaluating resource allocation efficiency); goal achievement (measuring the quality of the final result); and risk exposure (quantifying the impact of potential risks). These parameters and metrics together constitute the initial simulation configuration, providing a framework and standard for subsequent Monte Carlo simulations.

[0118] S9.2: Based on the initial simulation configuration, the system uses the Monte Carlo method to generate multiple sets of random disturbance and fluctuation scenarios. The core of the Monte Carlo method is to simulate the behavior of complex systems through random sampling, making it particularly suitable for handling problems involving uncertainty. The system first identifies key random variables in the path execution process, such as the execution time of each node, resource availability, and the probability of external events; then, it defines an appropriate probability distribution for each variable (such as a normal distribution, exponential distribution, or a custom empirical distribution); next, it randomly samples from these distributions to generate specific values ​​for each variable in each scenario; finally, it combines these values ​​to construct a complete simulation scenario. By repeating this process multiple times, the system generates a set of simulation scenarios containing hundreds or thousands of different scenarios, covering various variations that may be encountered during path execution, from common fluctuations to rare extreme events.

[0119] S9.3: Based on the generated simulation scenario set, the system performs simulation runs on dynamically updated personalized paths. For each simulation scenario, the system fully simulates the execution process of the path: executing each node according to the node order and dependencies defined in the path; determining the actual execution time and resource consumption of each node based on random parameters in the scenario; simulating the interaction and dependency effects between nodes; and recording various events and indicator changes during the execution process. During the simulation, the system collects detailed performance indicator data, including: the start and end times of each node; the time series of resource usage; the selection results of key decision points; and the achievement status of the final goal. These data are aggregated to form a simulation results dataset, providing comprehensive basic data for subsequent analysis.

[0120] S9.4: Based on the simulation results dataset, the system performs in-depth statistical analysis and risk assessment. Statistical analysis includes: calculating the expected value, variance, quantiles, and extreme values ​​of key indicators; identifying correlations and dependencies between indicators; and detecting abnormal patterns and failure scenarios. Risk assessment includes: identifying high-risk nodes (nodes with high failure probability or significant impact); quantifying the probability of occurrence and potential impact of various risks; and evaluating the overall robustness of the path and its sensitivity to random disturbances. Based on these analysis results, the system makes final adjustments to the dynamically updated personalized path. Possible adjustments include: increasing the buffer time of key nodes; optimizing resource allocation for high-risk nodes; adjusting dependencies between nodes to improve parallelism; or adding alternative solutions to address specific risk scenarios. Finally, the system outputs a fully validated and optimized executable path, along with detailed decision support information, including expected performance indicators, potential risk points, suggested coping strategies, and sensitivity analysis results, providing comprehensive decision support for actual execution.

[0121] In the medical field, taking the optimization of comprehensive treatment pathways for patients with complex diabetes as an example, the system ensures that personalized treatment pathways can cope with various changes and uncertainties in the actual medical environment through two key stages: dynamic adjustment and robustness verification, ultimately providing reliable treatment decision support for both doctors and patients.

[0122] When a hospital admits a diabetic patient, Mr. X, over 60 years old, with coronary heart disease, hypertension, and mild renal insufficiency, the system first begins a dynamic adjustment process based on the personalized treatment pathway plan generated in the previous stages. The system receives this personalized treatment pathway and simultaneously acquires Mr. X's real-time monitoring data (such as blood glucose levels, blood pressure changes, and renal function indicators) and feedback information (such as drug responses, discomfort symptoms, and assessment of daily activity abilities). For example, the system might detect significant fluctuations in Mr. X's blood glucose (peaking at 12 mmol / L after meals, and as low as 3.5 mmol / L at night), mild stomach discomfort with a certain hypoglycemic drug, and a desire to reduce hospital visits due to the distance from his residence. Based on this real-time data, the system constructs an initial state space, mapping the patient's multidimensional physiological indicators and treatment adherence to various nodes of the treatment pathway.

[0123] Next, the system applies a Bayesian network model to model the conditional probability relationships between nodes in the personalized treatment pathway. For example, by analyzing historical data, the system establishes a conditional probability chain of "postprandial blood glucose level → insulin dosage adjustment → risk of nocturnal hypoglycemia → frequency of monitoring the next day". When Mr. A's postprandial blood glucose reaches 11 mmol / L, the system calculates that the necessity of increasing insulin dosage is high (probability 0.82), but also predicts that this may lead to an increased risk of nocturnal hypoglycemia (probability 0.35). The system further calculates the execution probability and priority of each treatment node. For example, it finds that compared to the weekly outpatient follow-up arranged in the standard pathway, Mr. A needs to increase the frequency of remote blood glucose monitoring (execution necessity probability 0.91) and adjust the combination of oral hypoglycemic drugs (execution necessity probability 0.87), while some routine examinations can be appropriately postponed (execution necessity probability is only 0.42).

[0124] Building upon this foundation, the system incorporates reinforcement learning algorithms to learn the optimal path adjustment strategy. Specifically, the system treats changes in the patient's state as environmental feedback, various adjustments to the treatment path as the action space, and a comprehensive score of treatment effectiveness (such as stable blood glucose control) and patient experience (such as minimizing adverse reactions and maximizing convenience of medical care) as the reward function. By interacting with historical treatment data from a large number of similar patients within the hospital, the system learns the optimal adjustment strategy to be adopted under different states. For example, the system learns that when a patient experiences the combined symptoms of "postprandial hyperglycemia + nocturnal hypoglycemia + stomach discomfort," a better strategy than increasing the insulin dose is to adjust the type of oral medication and advance the administration time by 30 minutes.

[0125] Based on the established dynamic adjustment decision model, the system automatically adjusts Mr. X's treatment pathway by combining his latest monitoring data and feedback. For example, the system changed the originally planned weekly outpatient follow-up to once every two weeks plus daily remote blood glucose monitoring; replaced the previously scheduled standard dose of metformin with a sustained-release formulation and adjusted the administration time to 15 minutes before meals; inserted a new TCM adjunctive therapy node to alleviate stomach discomfort; removed the originally planned exercise intensity test and replaced it with gentler home exercise guidance; and adjusted the fundus examination time from the second week to the fourth week, but increased the frequency of self-vision monitoring. These adjustments form a dynamically updated personalized treatment pathway that is more in line with Mr. X's actual situation and needs.

[0126] After completing the dynamic adjustment, the system enters the robustness verification phase to ensure that the optimized path can cope with various possible changes and challenges. The system first sets the Monte Carlo simulation parameters, including the number of simulations (e.g., 1000), the time span (e.g., a 6-month treatment cycle), and key evaluation indicators (e.g., glycemic control achievement rate, complication risk index, quality of life score, and medical resource consumption).

[0127] Based on these configurations, the system generates multiple sets of random disturbances and fluctuation scenarios to simulate various situations that Mr. A might face during actual treatment. These scenarios include physiological fluctuations (such as blood pressure fluctuations due to seasonal changes, and abnormal blood sugar levels caused by irregular eating habits), behavioral changes (such as occasionally missing a medication dose, or temporarily changing activity plans), environmental factors (such as family emergencies, or transportation difficulties affecting medical treatment), and medical system factors (such as specific doctors taking leave, or examination delays due to equipment maintenance). For example, the system might simulate a complex disturbance scenario such as "irregular eating habits during the Spring Festival + missing two medication doses + a common cold," or a situation such as "high summer temperatures + mild dehydration + temporary inability to attend medical treatment on time."

[0128] In these simulation scenarios, the system simulates the operation of dynamically updated treatment pathways and collects the execution results and performance indicators for each scenario. For example, in the "irregular eating habits during the Spring Festival" scenario, the system found that the blood sugar control target achievement rate decreased by 15%, but by increasing the frequency of remote monitoring and temporarily adjusting the medication dosage, the risk of complications remained within an acceptable range. In the "high temperature in summer" scenario, the increased risk of dehydration led to fluctuations in kidney function indicators, requiring increased reminders for water intake and temporary adjustments to the dosage of antihypertensive medication.

[0129] Based on these simulation results, the system conducted statistical analysis and risk assessment. The analysis showed that in approximately 82% of the 1000 simulations, blood glucose control was good, moderate adjustments were needed in 15% of cases, and emergency medical intervention was required in 3% of cases. The system identified several potential risk points, such as: a significantly increased risk of worsening renal function when the average blood glucose level exceeds 10 mmol / L for three consecutive days; an increased risk of hypotension when antihypertensive drugs are adjusted simultaneously with certain hypoglycemic drugs; and a doubling risk of dehydration complications during hot summer months without increased monitoring of fluid intake.

[0130] In response to these identified risk points, the system makes final adjustments to the treatment pathway. For example, it adds an alert trigger mechanism for "three consecutive days of elevated blood sugar" to automatically schedule a follow-up kidney function test; it establishes a safety check process for synergistic adjustment of antihypertensive and hypoglycemic drugs; and it increases the frequency of water intake reminders and urine output monitoring during the hot summer months. The system also designs alternative plans for possible emergencies, such as alternative remote consultations for those unable to attend a scheduled appointment, or temporary dietary guidance for special circumstances.

[0131] Ultimately, the system outputs an optimized treatment pathway that has been dynamically adjusted and robustly validated, along with corresponding decision support information. This pathway is not only highly personalized, tailored to Mr. X's specific health condition and lifestyle, but also adaptable to various real-world disruptions. The medical team receives a detailed treatment plan, including a medication adjustment schedule, monitoring frequency arrangements, risk warning thresholds, and coping strategies. Mr. X receives easy-to-understand treatment guidance, including daily medication reminders, dietary recommendations, activity plans, and self-monitoring guidelines.

[0132] In practice, this optimized personalized treatment pathway brought significant improvement to Mr. Mou: after 6 months, his glycated hemoglobin dropped from the initial 8.5% to 6.8%, reaching the standard of good control; his blood pressure stabilized at around 135 / 85 mmHg; his renal function indicators did not deteriorate further; his quality of life score improved by 20%; compared with the standard treatment pathway, the number of hospital visits decreased by 40%, while the treatment effect improved.

[0133] The medical team also benefits: through the decision support information provided by the system, doctors can manage patients more efficiently, focusing more on key medical decisions while leaving routine monitoring and adjustments to the system. For example, when a patient's remote monitoring data shows an upward trend in blood sugar, the system automatically analyzes possible causes and provides adjustment suggestions. Doctors only need to confirm or fine-tune these suggestions, greatly improving work efficiency.

[0134] Through this dual guarantee of dynamic adjustment and robustness verification, personalized pathway optimization systems in the medical field not only improve the accuracy and efficiency of chronic disease management, but also enhance the adaptability of treatment plans to the complexity and uncertainty of the real world, providing patients and medical teams with reliable and easy-to-implement treatment decision support.

[0135] This embodiment provides a dynamic digital path generation system, including: a data preprocessing module for acquiring multidimensional heterogeneous data, preprocessing and extracting features from the multidimensional heterogeneous data to obtain a preliminary transformed feature matrix; a manifold dimensionality reduction module for constructing and reducing the dimensionality of the preliminary transformed feature matrix through nonlinear manifold representation to obtain a standardized feature vector set; a template matching module for constructing a mapping relationship between features and path templates through a random neural network based on the standardized feature vector set and a pre-built knowledge graph and standard path template library, and training the model to output an initial path template and the corresponding key node sequence; and a path optimization module. The system comprises four modules: a module for solving a personalized path scheme; a module for dynamic adjustment; and a module for simulation evaluation. The module is used to transform the path optimization problem into a source-to-target distribution transmission problem based on the initial path template and corresponding key node sequence, combined with user individual characteristics and resource constraints, using optimal transmission theory. The module is responsible for obtaining a personalized path scheme by combining the personalized path scheme with real-time monitoring data and feedback information, modeling the conditional probability relationships between nodes using Bayesian networks, and constructing a dynamic adjustment decision model using reinforcement learning algorithms. This automatically executes path node adjustment operations and outputs a dynamically updated personalized path.

[0136] The data preprocessing module is responsible for acquiring and processing multidimensional heterogeneous data, including basic parameters, state indicators, matching scores, and complexity quantification values. This module first performs data cleaning, including missing value imputation, outlier detection, and standardization. Then, it performs feature engineering for different types of features, such as one-hot encoding of categorical features, extraction of statistics from time-series features, and vectorization of text features. Finally, it outputs a pre-transformed feature matrix, laying the foundation for subsequent manifold representation and dimensionality reduction.

[0137] The manifold dimensionality reduction module receives the initially transformed feature matrix and constructs a nonlinear manifold representation using the Hamiltonian-Poisson integrator technique. This technique preserves the data topology by solving a specific system of partial differential equations, achieving a unified representation of heterogeneous data. The module then calculates the geodesic distance matrix between data points in the manifold embedding representation and applies a manifold learning algorithm for dimensionality reduction, reducing the high-dimensional feature space to a preset target dimension and outputting a standardized feature vector set. This provides a high-quality feature representation for subsequent path template matching.

[0138] The template matching module, based on a standardized feature vector set, combines a pre-built knowledge graph and a standard path template library to construct a mapping relationship between features and path templates. This module designs and initializes a stochastic neural network architecture containing randomization layers, hidden layers, attention layers, and an output layer. It is trained using an improved stochastic gradient descent algorithm and incorporates random field theory to handle correlations and uncertainties between features. Through forward propagation and Monte Carlo sampling, the module calculates the matching probability between feature vectors and each path template, selecting the optimal initial path template and its key node sequence.

[0139] The path optimization module constructs a constrained optimization problem based on the initial path template and key node sequence, combined with individual user characteristics and system resource constraints. This module innovatively applies optimal transport theory, transforming the path optimization problem into a transport problem from a source distribution to a target distribution. It solves this problem using convex relaxation methods and Kantorovich duality to obtain an optimized transport plan. Subsequently, the module reconstructs the path based on the transport plan and fine-tunes the node parameters, ultimately outputting a personalized path solution.

[0140] The dynamic adjustment module constructs an initial state space corresponding to the state and path based on the personalized path plan, combined with real-time monitoring data and feedback information. This module applies a Bayesian network to model the conditional probability relationships between nodes, calculates the execution probability and priority of each path node, and then introduces a reinforcement learning algorithm to learn the path adjustment strategy, constructing a dynamic adjustment decision model. Based on this model and real-time data, the module automatically performs insertion, deletion, replacement, and time adjustment operations on path nodes, outputting dynamically updated personalized paths, enabling the paths to be adaptive.

[0141] The simulation evaluation module, based on dynamically updated personalized paths, sets Monte Carlo simulation parameters and evaluation metrics. This module uses the Monte Carlo method to generate multiple sets of random disturbance and fluctuation scenarios, simulating path execution under different conditions and collecting path execution results and performance metrics for each scenario. Through statistical analysis and risk assessment, the module identifies potential risk points and makes final adjustments, outputting the final executable optimized path and corresponding decision support information to ensure the stability and reliability of the path.

[0142] This invention provides a method and system for dynamically generating digital paths, capable of processing high-dimensional heterogeneous data, dynamically adjusting, and finding the global optimal solution under multiple constraints. This method and system are particularly suitable for complex scenarios requiring strict process control and exhibiting significant individual differences, such as precision manufacturing, healthcare, and complex project management. It enables optimized resource allocation and refined process management, improving system efficiency and accuracy.

[0143] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these 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 present invention.

[0144] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the scope of protection of this application. Therefore, the patent protection scope of this application should be determined by the appended claims.

Claims

1. A method for dynamically generating digital paths, characterized in that, include: Acquire multidimensional heterogeneous data, preprocess and extract features from the multidimensional heterogeneous data to obtain a preliminary transformed feature matrix; Based on the feature matrix of the initial transformation, a standardized feature vector set is obtained through nonlinear manifold representation construction and dimensionality reduction. Based on the standardized feature vector set and the pre-built knowledge graph and standard path template library, a mapping relationship between features and path templates is constructed through a random neural network and the model is trained to output the initial path template and the corresponding key node sequence. Based on the initial path template and the corresponding key node sequence, combined with individual user characteristics and resource constraints, the path optimization problem is transformed into a transmission problem from source distribution to target distribution through optimal transmission theory, and then solved and optimized to obtain a personalized path scheme. Based on the personalized path scheme, combined with real-time monitoring data and feedback information, a dynamic adjustment decision model is constructed by modeling the conditional probability relationship between nodes through a Bayesian network and combining it with a reinforcement learning algorithm. The model automatically executes path node adjustment operations and outputs a dynamically updated personalized path. Based on the dynamically updated personalized path, Monte Carlo simulation technology is applied to generate a simulation scenario and run the path simulation. Adjustments are made through statistical analysis and risk assessment, and the final executable optimized path and corresponding decision support information are output.

2. The method according to claim 1, characterized in that, The preprocessing and feature extraction of the multidimensional heterogeneous data to obtain a preliminary transformed feature matrix includes: Receive multidimensional heterogeneous data, which includes basic parameters, state indicators, matching degree scores and complexity quantification values. Through missing value imputation, outlier detection and standardization, output a preprocessed multi-source heterogeneous dataset. Based on the preprocessed multi-source heterogeneous dataset, one-hot encoding is performed on categorical features, statistical extraction is performed on temporal features, and vectorization is performed on text features to output the feature matrix of the preliminary transformation.

3. The method according to claim 2, characterized in that, The feature matrix based on the preliminary transformation is used to construct and reduce dimensions through nonlinear manifold representation to obtain a standardized feature vector set, including: Based on the feature matrix of the preliminary transformation, the Hamiltonian-Poisson integrator technique is introduced to solve partial differential equations to maintain the data topology and output a unified manifold embedding representation. Based on the unified manifold embedding representation, the geodesic distance matrix between data points in the unified manifold embedding representation is calculated. The manifold learning algorithm is applied to perform dimensionality reduction processing based on the geodesic distance matrix to reduce the feature space dimension to a preset target dimension, and the standardized feature vector set is output.

4. The method according to claim 1, characterized in that, Based on the standardized feature vector set and the pre-built knowledge graph and standard path template library, the system constructs a mapping relationship between features and path templates through a random neural network and trains the model to output an initial path template and the corresponding key node sequence, including: Receive the standardized feature vector set, extract the pre-built knowledge graph and standard path template library from the knowledge base, establish the correspondence between the standardized feature vector set and the path template through association mapping, and output the training dataset; Based on the training dataset, a stochastic neural network architecture including a randomization layer, a hidden layer, an attention layer, and an output layer is constructed. The network parameters and learning rate are set, and the weights are randomly initialized. The initialized stochastic neural network model is output. Based on the initialized stochastic neural network model and the training dataset, the stochastic gradient descent algorithm is applied to train the model. Random field theory is introduced to construct a Markov random field model to represent the conditional independence structure between features, and the trained stochastic neural network model is output. Based on the trained stochastic neural network model, and combined with the standardized feature vector set, the matching probability between the standardized feature vector set and each standard path template in the standard path template library is calculated through forward propagation and Monte Carlo sampling. The path template with the highest matching probability value is selected, and the initial path template and the corresponding key node sequence are output.

5. The method according to claim 4, characterized in that, The process involves training the model using the stochastic gradient descent algorithm based on the initialized stochastic neural network model and the training dataset, introducing random field theory to construct a Markov random field model to represent the conditional independence structure between features, and outputting the trained stochastic neural network model, including: Based on the initialized stochastic neural network model and the training dataset, the model is trained using the Adam optimizer combined with a periodic learning rate adjustment strategy and gradient pruning technique to obtain the basic training model. Based on the aforementioned basic training model, a Markov random field model is constructed to represent the conditional independence structure between features, a conditional probability dependency graph between features is established, and the mutual influence between features is learned, outputting the trained random neural network model.

6. The method according to claim 5, characterized in that, The trained stochastic neural network model, combined with the standardized feature vector set, calculates the matching probability between the standardized feature vector set and each standard path template in the standard path template library through forward propagation and Monte Carlo sampling, selects the path template with the highest matching probability value, and outputs the initial path template and the corresponding key node sequence, including: Based on the trained stochastic neural network model and the standardized feature vector set, the number of Monte Carlo samplings is set to a preset threshold. Forward propagation calculations are performed for the preset threshold number of times. The average of all sampling results is taken to obtain a stable matching probability distribution between the standardized feature vector set and each standard path template in the standard path template library. Forward propagation calculations are performed multiple times using Monte Carlo sampling. The average of all sampling results is taken to obtain a stable matching probability distribution. Based on the stable matching probability distribution, the standard path templates in the standard path template library are sorted from high to low according to the matching probability value. The top K standard path templates with the highest matching probability are retained as the candidate standard path template set. Each standard path template in the candidate standard path template set is evaluated, and the standard path template corresponding to the maximum comprehensive score is selected. The initial path template and the corresponding key node sequence are then output.

7. The method according to claim 1, characterized in that, The process involves transforming the path optimization problem into a source-to-target distribution transmission problem using optimal transmission theory, based on the initial path template and corresponding key node sequence, combined with user individual characteristics and resource constraints, and solving and optimizing the solution to obtain a personalized path scheme, including: Receive the initial path template and the corresponding key node sequence, obtain user individual characteristics and system resource constraints, construct an optimization problem including objective function, decision variables and constraints, and output the constrained optimization problem definition; Based on the definition of the constrained optimization problem, a source distribution representing the initial path template and a target distribution representing the ideal personalized path are constructed. The transmission cost matrix between any two points in the state space is calculated and the transmission plan is initialized. The initial solution of the transmission plan is output. Based on the initial solution of the transmission plan, the optimal transmission problem is transformed into a Kantorovich dual form by applying the convex relaxation method. The dual problem is solved iteratively and the original transmission plan is reconstructed from the dual solution to output the optimized path scheme. Based on the optimized path scheme, a sensitivity analysis model for node parameter adjustment is constructed to optimize the execution time, resource allocation, and dependencies of nodes in the optimized path scheme, and output the personalized path scheme.

8. The method according to claim 7, characterized in that, The process involves transforming the optimal transmission problem into a Kantorovich dual form using a convex relaxation method based on the initial solution of the transmission plan. This is followed by iteratively solving the dual problem and reconstructing the original transmission plan from the dual solution, outputting an optimized path scheme. Based on the initial solution of the transmission plan, the Kantorovich dual variables are initialized, and the problem is solved iteratively using problem decomposition and parallel computing techniques to obtain the final dual solution. Based on the final dual solution, the original transmission plan is reconstructed and the modified path node parameters are generated, and the optimized path scheme is output.

9. The method according to claim 1, characterized in that, The process involves, based on the personalized path scheme and incorporating real-time monitoring data and feedback information, modeling the conditional probability relationships between nodes using a Bayesian network and constructing a dynamic adjustment decision model using a reinforcement learning algorithm. This model automatically executes path node adjustment operations and outputs a dynamically updated personalized path, including: Receive the personalized path scheme, obtain real-time monitoring data and feedback information, and output the initial state space corresponding to the state-path; Based on the initial state space, a Bayesian network is applied to model the conditional probability relationship between nodes in the personalized path scheme, calculate the execution probability and priority of each path node, and output the node probability distribution model. Based on the node probability distribution model, a reinforcement learning algorithm is introduced to learn the path adjustment strategy and output a dynamic adjustment decision model. Based on the dynamic adjustment decision model, combined with the real-time monitoring data and feedback information, the system automatically performs insertion, deletion, replacement, and time adjustment operations on path nodes in the personalized path scheme, and outputs the dynamically updated personalized path.

10. The method according to claim 1, characterized in that, The personalized path, based on the dynamically updated model, uses Monte Carlo simulation technology to generate a simulation scenario and run the path simulation. Adjustments are made through statistical analysis and risk assessment, outputting the final executable optimized path and corresponding decision support information, including: Receive the dynamically updated personalized path, set the Monte Carlo simulation parameters and evaluation metrics, and output the initial simulation configuration; Based on the initial simulation configuration, the Monte Carlo method is applied to generate multiple sets of random disturbance and fluctuation scenarios, and the simulation scenario set is output. Based on the simulation scenario set, the dynamically updated personalized path is simulated and run, and the path execution results and performance indicators under each scenario are collected, and the simulation result dataset is output. Based on the simulation result dataset, statistical analysis and risk assessment are performed to identify potential risk points and adjust the dynamically updated personalized path, outputting the final executable optimized path and corresponding decision support information.

11. A digital path dynamic generation system, characterized in that, include: The data preprocessing module is used to acquire multidimensional heterogeneous data, preprocess the multidimensional heterogeneous data and extract features to obtain a preliminary transformed feature matrix. The manifold dimensionality reduction module is used to construct and reduce the dimensionality of the feature matrix based on the initial transformation through nonlinear manifold representation to obtain a standardized feature vector set; The template matching module is used to construct the mapping relationship between features and path templates based on the standardized feature vector set and the pre-built knowledge graph and standard path template library, and to train the model by using a random neural network to output the initial path template and the corresponding key node sequence. The path optimization module is used to transform the path optimization problem into a transmission problem from the source distribution to the target distribution based on the initial path template and the corresponding key node sequence, combined with the user's individual characteristics and resource constraints, and to solve and optimize the problem to obtain a personalized path scheme. The dynamic adjustment module is used to automatically execute path node adjustment operations and output dynamically updated personalized paths based on the personalized path scheme, combined with real-time monitoring data and feedback information, by modeling the conditional probability relationship between nodes through a Bayesian network and constructing a dynamic adjustment decision model by combining reinforcement learning algorithms. The simulation evaluation module is used to generate simulation scenarios and run path simulations based on the dynamically updated personalized paths using Monte Carlo simulation technology. Adjustments are made through statistical analysis and risk assessment, and the final executable optimized path and corresponding decision support information are output.

Images