Multi-level process node prediction method and device, equipment and storage medium

Abstract

The invention provides a multi-level process node prediction method and device, equipment and a storage medium, and relates to the technical field of process management.The method comprises the steps that firstly, operation data of all levels in a target level process is obtained, an initial data set is constructed, the comprehensiveness and accuracy of model input are ensured, and a data foundation is laid for subsequent analysis; secondly, constructing a multi-level process comprehensive matrix based on the data set, systematically quantifying the incidence relation among the levels, breaking through the limitation of traditional single-level analysis, and remarkably improving the understanding ability of a complex process structure; then, through designing an improved self-evolution network architecture and integrating a game theory optimization strategy, the model realizes dynamic game and adaptive adjustment of parameters in a hidden layer, and a prediction strategy is continuously optimized in an iteration process; and finally, a stable and convergent fusion model is obtained after multiple rounds of iteration, the evolution law of a multi-level process can be accurately captured, and panoramic prediction of a future path is realized.

Description

Technical Field

[0001] The present application relates to the field of process management technology, and in particular to a multi-level process node prediction method, apparatus, device and storage medium. Background Art

[0002] As modern businesses continue to grow in complexity and diversity, multi-level process management and monitoring play a crucial role across various industries. The core goal of process management is to ensure efficient and coordinated operation of all links, promptly identify and resolve potential bottlenecks, and thus ensure the smooth and stable operation of the overall process.

[0003] However, existing process management and monitoring technologies still have many shortcomings when faced with complex and ever-changing business scenarios. In existing multi-level process management and monitoring models, since the monitoring structures adopted mostly tend to record single links or local nodes, it is impossible to more finely quantify and track the connection and resource sharing relationships between multi-layer processes in complex application environments. Even if some technical solutions can identify hierarchical coupling and implement association identification of cross-process nodes in a pre-set adjacency model, it is often difficult to quickly capture resource conflicts caused by sudden changes in process branches due to the lack of dynamic feedback on node business attributes and interaction intensity. Summary of the Invention

[0004] To address the above issues, this application provides a multi-level process node prediction method, which includes the following contents:

[0005] In a first aspect, the present application provides a multi-level process node prediction method, the method comprising:

[0006] Obtaining operation data of each level process in the target level process, and obtaining an initial data set based on the operation data of each level process;

[0007] Constructing a multi-level process comprehensive matrix based on the initial data set, wherein the multi-level process comprehensive matrix is ​​used to represent the relationship between the processes at each level;

[0008] Constructing an improved self-evolving network architecture based on the multi-level process integration matrix;

[0009] Incorporating a game theory optimization strategy into the hidden layer of the improved self-evolving network;

[0010] The improved self-evolving network is iterated based on the benefit feedback of the game theory optimization strategy; a stably converged fusion model is obtained after multiple rounds of iteration, and the fusion model is used to perform a panoramic prediction of the future path of the multi-level process.

[0011] Optionally, the acquiring of the operating data of each level of the target level process and obtaining the initial data set based on the operating data of each level of the target level process includes:

[0012] For the parent process, child process, and grandchild process, four monitoring indicators are collected: execution time, waiting queue length, external system call delay, and number of retries;

[0013] The monitoring indicators are recorded as a multidimensional sequence in chronological order, and weight factors of the parent process, child process and grandchild process are introduced to calculate the weighted monitoring value by combining cumulative summation with exponential decay;

[0014] The threshold strategy is used to mark abnormal points and remove duplicates of the monitoring values ​​that deviate from the mean, and then the obtained monitoring data matrix is ​​normalized to obtain an initial data set with a unified value range.

[0015] Optionally, constructing a multi-level process comprehensive matrix based on the initial data set includes:

[0016] Selecting multiple process links of the parent process, child process and grandchild process as a node set;

[0017] Combining the order and resource dependency in the node set, an adjacency matrix is ​​defined for each layer of the process. Then, a cross-layer coupling matrix and a layer weight factor are introduced to perform weighted fusion on the adjacency relationships among the parent process, child process, and grandchild process to obtain a multi-level process comprehensive matrix.

[0018] Preset key indicators are added to each node in the multi-level process comprehensive matrix to form a node feature vector.

[0019] Optionally, constructing an improved self-evolving network architecture based on the multi-level process comprehensive matrix includes:

[0020] In a preset nested network template, receiving input including node feature vectors and comprehensive matrices;

[0021] The prediction error and network structure complexity are measured using a piecewise defined adaptive network cost function;

[0022] When the cost function exceeds the set threshold, the network depth or parameter scale is dynamically expanded to improve the fitting ability. When the cost function is at a low level, the network depth is reduced accordingly to improve the operating efficiency.

[0023] Optionally, the incorporation of a game theory optimization strategy into the hidden layer of the improved self-evolving network includes:

[0024] Treat each node in the parent process, child process and grandchild process as the game subject participating in resource competition or collaboration;

[0025] Construct a profit function for each node based on resource occupancy and service priority information;

[0026] The profit signal obtained by the profit function is used as an additional input to correct the hidden layer output, so as to enhance the network's ability to focus on and identify potential bottleneck nodes.

[0027] Optionally, the iterating the improved self-evolving network based on the benefit feedback of the game theory optimization strategy includes:

[0028] Define a deviation metric function and track the degree of deviation between the forecast error and the return distribution;

[0029] When the cumulative deviation is significant, the activation function or connection parameters of the hidden layer are adjusted by gain or attenuation, so that the network can take into account both global prediction performance and precise focus on key nodes.

[0030] Optionally, the stable convergence fusion model is used to output the congestion risk value of each node at a future moment and the impact range of the execution time of the entire process, thereby locating the root causes that may occur in multi-level processes.

[0031] In a second aspect, the present application provides a multi-level process node prediction device, which includes:

[0032] an acquisition unit, configured to acquire operation data of each level of the target level process, and obtain an initial data set based on the operation data of each level of the process;

[0033] A first construction unit is configured to construct a multi-level process comprehensive matrix based on the initial data set, wherein the multi-level process comprehensive matrix is ​​configured to represent the relationship between the processes at each level;

[0034] A second construction unit is configured to construct an improved self-evolving network architecture based on the multi-level process comprehensive matrix; and to incorporate a game theory optimization strategy into the hidden layer of the improved self-evolving network;

[0035] an iterative optimization unit, configured to iterate the improved self-evolving network based on the benefit feedback of the game theory optimization strategy, and obtain a stable convergent fusion model after multiple rounds of iterations;

[0036] The prediction unit is used to use the stable convergence fusion model to perform a panoramic prediction of the future path of the multi-level process.

[0037] Optionally, the acquisition unit is specifically used to collect four monitoring indicators, namely, execution time, waiting queue length, external system call delay, and number of retries, for the parent process, child process, and grandchild process respectively;

[0038] The monitoring indicators are recorded as a multidimensional sequence in chronological order, and weight factors of the parent process, child process and grandchild process are introduced to calculate the weighted monitoring value by combining cumulative summation with exponential decay;

[0039] The threshold strategy is used to mark abnormal points and remove duplicates of the monitoring values ​​that deviate from the mean, and then the obtained monitoring data matrix is ​​normalized to obtain an initial data set with a unified value range.

[0040] Optionally, the first building unit is specifically used to:

[0041] Selecting multiple process links of the parent process, child process and grandchild process as a node set;

[0042] Combining the order and resource dependency in the node set, an adjacency matrix is ​​defined for each layer of the process. Then, a cross-layer coupling matrix and a layer weight factor are introduced to perform weighted fusion on the adjacency relationships among the parent process, child process, and grandchild process to obtain a multi-level process comprehensive matrix.

[0043] Preset key indicators are added to each node in the multi-level process comprehensive matrix to form a node feature vector.

[0044] Optionally, the second construction unit is configured to construct an improved self-evolving network architecture based on the multi-level process comprehensive matrix, including:

[0045] In a preset nested network template, receiving input including node feature vectors and comprehensive matrices;

[0046] The prediction error and network structure complexity are measured using a piecewise defined adaptive network cost function;

[0047] When the cost function exceeds the set threshold, the network depth or parameter scale is dynamically expanded to improve the fitting ability. When the cost function is at a low level, the network depth is reduced accordingly to improve the operating efficiency.

[0048] Optionally, the second construction unit integrates a game theory optimization strategy into the hidden layer of the improved self-evolving network, including:

[0049] Treat each node in the parent process, child process and grandchild process as the game subject participating in resource competition or collaboration;

[0050] Construct a profit function for each node based on resource occupancy and service priority information;

[0051] The profit signal obtained by the profit function is used as an additional input to correct the hidden layer output, so as to enhance the network's ability to focus on and identify potential bottleneck nodes.

[0052] Optionally, the iterative optimization unit is specifically used to:

[0053] Define a deviation metric function and track the degree of deviation between the forecast error and the return distribution;

[0054] When the cumulative deviation is significant, the activation function or connection parameters of the hidden layer are adjusted by gain or attenuation, so that the network can take into account both global prediction performance and precise focus on key nodes.

[0055] Optionally, the stable convergence fusion model is used to output the congestion risk value of each node at a future moment and the impact range of the execution time of the entire process, thereby locating the root causes that may occur in multi-level processes.

[0056] In a third aspect, the present application provides a device comprising a memory and a processor, wherein the memory is used to store instructions or codes, and the processor is used to execute the instructions or codes so that the device executes the multi-level process node prediction method introduced in any implementation of the first aspect.

[0057] In a fourth aspect, the present application provides a computer-readable storage medium storing a code. When the code is executed, the device executing the code implements the multi-level process node prediction method described in any implementation of the first aspect.

[0058] This application provides a multi-level process node prediction method. When executing this method, first, by obtaining the operating data of each level in the target level process and constructing an initial data set, the comprehensiveness and accuracy of the model input are ensured, laying a data foundation for subsequent analysis; secondly, based on the data set, a multi-level process comprehensive matrix is ​​constructed to systematically quantify the correlation between each level, breaking through the limitations of traditional single-level analysis, and significantly improving the ability to understand complex process structures; then, by designing an improved self-evolutionary network architecture and integrating game theory optimization strategies, the model realizes dynamic game and adaptive adjustment of parameters in the hidden layer, and continuously optimizes the prediction strategy during the iteration process; finally, after multiple rounds of iteration, a stable convergence fusion model is obtained, which can accurately capture the evolution law of multi-level processes and realize a panoramic prediction of future paths.

[0059] In summary, this method achieves efficient prediction and management of complex, multi-level processes by constructing a comprehensive multi-level process matrix, employing an improved self-evolving network architecture, and incorporating game-theoretic optimization strategies. This method not only comprehensively and meticulously represents the complex relationships between processes at all levels and captures resource conflicts caused by changes in process branches, but also dynamically adjusts the network structure based on real-time data, balancing prediction accuracy and operational efficiency. Game-theoretic optimization strategies enhance the ability to identify bottleneck nodes and further improve the ability to predict potential congestion. Ultimately, the fusion model converges stably and outputs a comprehensive prediction result, quantifying the congestion probability of each node and the extent of its impact on the overall process, providing managers with comprehensive and forward-looking decision support. This approach significantly improves the efficiency and reliability of multi-level process management, optimizes resource allocation, and enhances the operational efficiency and competitiveness of enterprises in complex business environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in this embodiment or the prior art, the following briefly introduces the drawings required for use in the embodiment or the prior art description. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0061] Figure 1 A flowchart of a multi-level process node prediction method provided in an embodiment of the present application;

[0062] Figure 2 A flow chart of a method for generating an initial data set provided in an embodiment of the present application;

[0063] Figure 3 A flowchart of a method for constructing a multi-level process integration matrix provided in an embodiment of the present application;

[0064] Figure 4 A flowchart of a method for constructing an improved self-evolving network architecture based on the multi-level process comprehensive matrix provided in an embodiment of the present application;

[0065] Figure 5 A flowchart of a method for integrating a game theory optimization strategy into the hidden layer of the improved self-evolving network provided in an embodiment of the present application;

[0066] Figure 6 A flowchart of iterating the improved self-evolving network based on the profit feedback of the game theory optimization strategy provided in an embodiment of the present application;

[0067] Figure 7 A flowchart of an embodiment of the present application for predicting the future path of a multi-level process using a fusion model;

[0068] Figure 8 A schematic diagram of the structure of a multi-level process node prediction device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0069] In order to make the purpose, technical solutions and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0070] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions.

[0071] Figure 1 A flowchart of a multi-level process node prediction method provided in an embodiment of the present application. Figure 1 As shown, the multi-level process node prediction method provided by the embodiment of the present application may include:

[0072] S101: Acquire the operation data of each level process in the target level process, and obtain an initial data set based on the operation data of each level process.

[0073] The target-level process refers to the multi-level, structured business execution path involved in enterprise operations or complex systems, which includes multiple nested levels of processes. In the embodiment of the present application, the target-level process is clearly divided into three levels: parent process, child process, and grandchild process. The parent process represents the overall business goal or task framework, the child process is a detailed decomposition of the parent process, and the grandchild process is a further split of the child process. The three together constitute a hierarchical process system.

[0074] To fully demonstrate the operational status of processes at all levels, key monitoring indicators must be collected for each level, including execution time, queue length, external system call latency, and number of retries. Execution time refers to the time span from process initiation to completion, reflecting the process's operational efficiency; queue length refers to the number of pending tasks in the process, reflecting the process's load; external system call latency refers to the response wait time when calling external interfaces or services within the process, measuring the process's sensitivity to external dependencies; and retries refers to the number of times a process is re-executed due to exceptions or failures, reflecting the process's stability and fault tolerance.

[0075] Collected monitoring metrics must be recorded in chronological order as a multidimensional sequence—a collection of vectors indexed by timestamps—to preserve the data's temporal nature. To quantify the relative importance of processes at each level, weighting factors are introduced, assigning different weights to parent processes, child processes, and grandchild processes. By combining cumulative summation with exponential decay, weighted monitoring values ​​are calculated for each level of the process, comprehensively reflecting both long-term trends and short-term fluctuations.

[0076] In order to eliminate outliers in the data, a threshold strategy is used to mark outliers and remove duplicates from the monitoring values ​​that deviate from the mean. Specifically, by calculating the mean and standard deviation of the monitoring values, data that deviates from the mean by more than a set threshold, such as 3 times the standard deviation, is marked as an outlier and removed. At the same time, duplicated data is deduplicated to avoid data redundancy.

[0077] Finally, the preprocessed monitoring data matrix is ​​normalized, linearly mapping the values ​​of each indicator to a uniform interval to eliminate the impact of dimensional differences on subsequent analysis. The normalized data set constitutes the initial data set, providing standardized input for the subsequent construction of the multi-level process matrix and prediction model.

[0078] Through the above steps, the systematic collection, cleaning and standardization of multi-level process operation data are achieved, laying a solid data foundation for subsequent process modeling and predictive analysis.

[0079] S102: Construct a multi-level process comprehensive matrix based on the initial data set, wherein the multi-level process comprehensive matrix is ​​used to represent the relationship between the processes at each level.

[0080] First, for each of the parent process, child process, and grandchild process, multiple process steps are selected as node sets. A node set is the collection of key steps in each level of the process. These node sets at different levels are then merged together to form a single node set, which is then used to construct the comprehensive matrix.

[0081] Taking into account the execution order and resource dependencies of processes, an adjacency matrix is ​​defined for each process layer. This matrix represents the connectivity between nodes within that layer, specifically which nodes are adjacent and the execution order or dependencies between them. A cross-layer coupling matrix is ​​introduced to describe the relationships between processes at different levels, specifically how parent processes, child processes, and grandchild processes influence each other. Furthermore, a layer weight factor is assigned to each process layer to indicate its importance or influence within the overall multi-layer process.

[0082] By weightedly integrating the adjacency relationships between parent, child, and grandchild processes, a comprehensive matrix representing the multi-level process structure is generated. This matrix comprehensively reflects the inherent connections and overall structure between processes at each level. Based on this comprehensive representation matrix, key business context indicators such as execution time and resource consumption are loaded for each node to form a node feature vector. This node feature vector comprehensively describes the node's attributes and status, providing a foundation for subsequent analysis and prediction.

[0083] S103: Constructing an improved self-evolving network architecture based on the multi-level process comprehensive matrix.

[0084] Based on the previously obtained multi-level process synthesis matrix, an improved self-evolving network architecture is constructed. Specifically, using a pre-defined nested network template, data containing node feature vectors and a synthesis representation matrix is ​​input into this network. Next, a piecewise adaptive network cost function is used to measure prediction error and network complexity. Based on the performance of the cost function, the network structure is dynamically adjusted to achieve better performance and efficiency.

[0085] S104: Incorporating a game theory optimization strategy into the hidden layer of the improved self-evolving network.

[0086] Each node in the parent, child, and grandchild processes is considered a player in the game, competing for resources or collaborating. A payoff function is constructed for each node based on information such as the proportion of resources it occupies during operation and the importance of the business it handles. This payoff function calculates each node's payoff in the game. The resulting payoff signal is then used as additional input to modify the output of the hidden layer. In this way, the network can more keenly focus on and identify potential bottleneck nodes—those that could hinder the efficient operation of the entire process.

[0087] S105 , iterating the improved self-evolving network based on the benefit feedback of the game theory optimization strategy.

[0088] Reward feedback refers to feeding reward information derived from a reward function back into the network's subsequent processing, guiding adjustments to network parameters and structure. In improved self-evolving networks, the iterative process involves the network continuously optimizing performance through a series of operations, including forward and backward propagation, based on input data, the current network structure, and parameters. Each iteration adjusts the network toward a more optimal path, ensuring a better fit to the data and task performance.

[0089] The iterative process of the improved self-evolving network based on the reward feedback of the game theory optimization strategy specifically includes first defining a deviation metric function, which is used to measure the difference between the network's prediction error and the reward distribution. By calculating the difference between the prediction error and relevant indicators of the reward distribution, such as the mean and variance, a deviation degree value is obtained to determine whether the network has experienced abnormal deviation. The deviation degree function is then used to track the degree of deviation between the network's prediction error and the reward distribution. When the deviation degree calculated by the deviation metric function accumulates and reaches a high threshold, the cumulative deviation is considered significant. This means that the difference between the network's prediction error and the reward distribution has exceeded the normal range, which may indicate that there are some problems with the network and that adjustments and optimizations, such as gain or attenuation control, are required. This allows the network to ensure overall prediction performance while accurately focusing on key nodes.

[0090] S106. After multiple rounds of iterations, a stable and convergent fusion model is obtained, and the fusion model is used to perform a panoramic prediction of the future path of the multi-level process.

[0091] After multiple rounds of iterative training, a stable and convergent fusion model was ultimately obtained. This fusion model incorporates the previously mentioned self-evolving network architecture and game-theoretic optimization strategies. This fusion model enables comprehensive and detailed predictions of the future trajectory of multi-level processes—a panoramic forecast. It also identifies potential congestion risks within the process. Specifically, when the network reaches convergence, the model quantifies the congestion probability of each node in the parent, child, and grandchild processes, as well as the potential downstream queuing delays. The model then outputs the congestion risk value for each node at a future time, as well as the impact range on the overall process execution time. This information can be used to identify key nodes or factors within the multi-level process that cause process anomalies or inefficiencies.

[0092] The above embodiment introduces the multi-level process node prediction method in this application. In order to present the technical details and implementation logic of this solution more clearly, comprehensively and in-depth, the above process is introduced in more detail through multiple specific embodiments.

[0093] In one implementation of the embodiment of the present application, the specific process of obtaining the operating data of each level process in the target level process and obtaining the initial data set based on the operating data of each level process is as follows: Figure 2 As shown, Figure 2 A flow chart of a method for generating an initial data set provided in an embodiment of the present application:

[0094] S201. For the parent process, child process and grandchild process, four monitoring indicators are collected respectively: execution time, waiting queue length, external system call delay and number of retries.

[0095] The above monitoring indicators are recorded as a multidimensional sequence in chronological order. Let the monitoring sequence of the parent process be recorded as , the monitoring sequence of the sub-process is recorded as , the monitoring sequence of the grandson process is recorded as ,in Represents the time index, Respectively expressed in Monitoring data vector collected at all times.

[0096] S202. Record the monitoring indicators as a multi-dimensional sequence in chronological order, introduce weight factors of the parent process, child process and grandchild process, and calculate the weighted monitoring value by combining cumulative summation with exponential decay.

[0097] make Represents the weight factors of the parent process, child process, and grandchild process respectively.

[0098] By combining cumulative summation with exponential decay, the following formula is defined as the calculation formula for weighted monitoring value, which is used to fuse historical data and real-time data in the initial stage:

[0099] ;

[0100] in, Represents the attenuation coefficient, which is used to control the data closer to the current moment to have a higher proportion in the weighted monitoring value. Indicates the Layer process in time monitoring data, .

[0101] S203: Using a threshold strategy, outliers are marked for monitoring values ​​that deviate from the mean and duplicates are removed. Then, the obtained monitoring data matrix is ​​normalized to obtain an initial data set with a uniform value range.

[0102] From the weighted monitoring value Starting from, combined with the monitoring sequence of each layer of process Perform outlier detection. Apply a threshold strategy to each data point and mark the monitored values ​​that deviate greatly from the mean as outliers. Then, perform deduplication on completely duplicated records within the same time index. Normalize the monitoring data matrix obtained after outlier detection and deduplication to ensure that the numerical distribution of different indicators or different levels of processes is in the same range. The normalization method is implemented by the following formula: and the corresponding monitoring vector Perform component-level mapping, letting Represents the normalized monitoring vector, then:

[0103]

[0104] in, represents the original monitoring vector after denoising and deduplication, Represent their global minimum and maximum values ​​respectively. After normalization, the value ranges of all monitoring indicators are uniformly mapped to , thus forming the initial data set.

[0105] In one implementation of the embodiment of the present application, the specific process of constructing a multi-level process comprehensive matrix based on the initial data set is as follows: Figure 3 As shown, Figure 3 A flowchart of a method for constructing a multi-level process comprehensive matrix provided in an embodiment of the present application:

[0106] S301: Select multiple process links of the parent process, child process and grandchild process as a node set.

[0107] For the parent process, child process and grandchild process, multiple process links are selected as node sets, recorded as as well as , representing the link node sets of the parent process, child process and grandchild process respectively.

[0108] Will Record it as the set of all nodes in the system, and identify the connection between each node according to the execution order and resource dependency recorded in the initial data set to form a multi-level node list.

[0109] S302. Based on the order and resource dependency in the node set, an adjacency matrix is ​​defined for each layer of the process. Then, a cross-layer coupling matrix and a layer weight factor are introduced to perform weighted fusion on the adjacency relationships among the parent process, child process, and grandchild process to obtain a multi-level process comprehensive matrix.

[0110] Based on the dependency order and coupling characteristics between nodes shown in the initial data set, the adjacency matrix is ​​defined for the parent process, child process and grandchild process respectively. 、 、 ,in correspond The node connection relationship in Corresponding to the node connection relationship in N2, Corresponding to the node connection relationship in N3.

[0111] make Represent the layer weight factors of the parent process, child process, and grandchild process respectively, and introduce the cross-layer coupling matrix , the following formula is used to perform weighted fusion on the multi-layer adjacency matrix to obtain the comprehensive matrix :

[0112] ;

[0113] in, Represents component-by-element multiplication operation, matrix Used to describe the resource dependency and linkage relationship between cross-layer nodes. represents the adjacency relationship in layer l, .

[0114] S303: Add preset key indicators to each node in the multi-level process comprehensive matrix to form a node feature vector.

[0115] based on The node association degree reflected in the above equation is added to each node, and the key indicators in the business context are added to each node, which is recorded as , respectively represent nodes The execution time, external dependency times, and business context allocation characteristic coefficients counted in the initial data set. The eigenvector of , use the following formula for weighted aggregation:

[0116] ;

[0117] in, and Represent the nonlinear amplification coefficients of external dependency and business context coefficients respectively, Represents the root node (parent process) and The global mapping ratio between nodes, Represents the associated business scenario tags extracted from the initial data set.

[0118] In one implementation of the embodiment of the present application, the specific process of constructing the improved self-evolving network architecture based on the multi-level process comprehensive matrix is ​​as follows: Figure 4 As shown, Figure 4A flowchart of a method for constructing an improved self-evolving network architecture based on the multi-level process comprehensive matrix provided in an embodiment of the present application:

[0119] S401: In a preset nested network template, receive input including node feature vectors and a comprehensive matrix.

[0120] The multi-layer progressive process representation matrix generated in the above embodiment and node feature vector set , select the initial network depth in the preset nested network template With the basic parameter set The initial network structure includes several hidden layers and corresponding connection weights, which are used to output the predicted values ​​of multi-layer process features when receiving node feature vectors.

[0121] S402. Utilize the segment-defined adaptive network cost function to measure the prediction error and the network structure complexity.

[0122] Based on the initial network, in order to measure the balance between prediction error and network complexity, an adaptive network cost function is defined .make Indicates the process nodes at time The prediction error, Represents the weighted coefficient corresponding to the criticality of the process link, Indicates time The complexity of the network structure, represents the constraint factor on complexity, Denotes the complexity threshold. The following formula is constructed as a piecewise cost function to describe the strategy of balancing error and structural scale when the complexity is low and high:

[0123] ;

[0124] Where K represents the total number of process nodes, Represents the prediction error term after multi-node weighting, and They respectively reflect the strategies of suppressing or amplifying network complexity.

[0125] when Not exceeding the threshold When , the cost function introduces the denominator to suppress the error from growing sharply in the global range; when Exceeding the threshold When , the additional parameter burden brought by large-scale networks is punished through multiplicative amplification.

[0126] S403. When the cost function exceeds the set threshold, the network depth or parameter scale is dynamically expanded to improve the fitting ability. When the cost function is at a low level, the network depth is correspondingly reduced to improve the operating efficiency.

[0127] Based on the adaptive control strategy of C(t), the network structure is dynamically expanded or reduced, and δ is the judgment threshold of the network cost function. At each time step t, the self-evolution adjustment of the network structure is achieved according to the following formula:

[0128] ;

[0129] in Indicates that at time step Update results of network depth or parameter scale.

[0130] When observed Greater than threshold , it means that the comprehensive performance under the current network structure cannot meet the prediction and complexity requirements, and it is necessary to add additional hidden layers or parameter units to improve the fitting ability.

[0131] when At or below When the network is updated, the network depth is reduced to avoid over-modeling and improve operational efficiency. This evolutionary update mechanism matches the multi-layer progressive process representation and can automatically expand or shrink the network size when new process branches are introduced or when real-time load mutations occur.

[0132] In one implementation of the embodiment of the present application, the specific process of integrating the game theory optimization strategy into the hidden layer of the improved self-evolving network is as follows: Figure 5 As shown, Figure 5 A flowchart of a method for integrating a game theory optimization strategy into the hidden layer of the improved self-evolving network provided in an embodiment of the present application, the process includes:

[0133] S501. Each node in the parent process, child process, and grandchild process is regarded as a game subject participating in resource competition or collaboration.

[0134] Based on the network structure and adaptive parameters in the above embodiment, the resource requirements and weighted relationships of different nodes in the parent, child, and grandchild process links are analyzed, and the node set is recorded as .Will Each node in the game is considered as a subject participating in the game, and the key attributes corresponding to the node are extracted, including the feature vector and the resource allocation or load evaluation parameters obtained from the prediction results of the improved self-evolving network , forming a node game environment. With the help of multi-layer adjacency relationships and the structural information of the adaptive network, global aggregation of process links at the parent-child level and even the grandchild level is achieved.

[0135] S502: Construct a profit function for each node based on resource occupancy and service priority information.

[0136] For each node State benefits in business scenarios, by comprehensively considering resource competition and dependent collaboration to build a benefit function .

[0137] make Representation node The priority weight of the process activity, Representation node At the moment The resource utilization rate of and risk penalty . Constructed by the following formula The fraction is combined with the power function exponential form:

[0138] ;

[0139] in, and Represent the power exponents of the collaborative benefit and risk terms, respectively, and are used to regulate the incentives or penalties for nodes with high resource utilization. Indicates the exponential amplification factor, which is used to amplify the resource allocation or load assessment parameters when the node resource is sufficient. Below threshold The penalty effect on node income is Represents the risk amplification weight.

[0140] In the denominator Represents exponential operation, when With threshold Larger gaps increase the attenuation of benefits dramatically, ensuring tighter constraints when key resources are scarce.

[0141] S503: Using the profit signal obtained by the profit function as an additional input to modify the hidden layer output, so as to enhance the network's ability to focus on and identify potential bottleneck nodes.

[0142] In the hidden layer structure of the self-evolving network, a revenue signal input module is added. Representation node At the original output of the hidden layer, the income signal is converted into As additional input and correction , forming the updated hidden layer output :

[0143] ;

[0144] in, Represents the weight adjustment coefficient of the income signal on the hidden layer output. By introducing Game optimization can further emphasize the moment It selects nodes that perform well in resource competition or collaborative relationships, guides network parameters to tilt towards potential blocking points or high-latency nodes during training, and strengthens the ability to predict and identify bottlenecks in key business processes.

[0145] The above steps S501 to S503 constitute a progressive process from building a business game environment to designing a node benefit function, and then to embedding the game signal into the hidden layer output. By integrating game theory ideas into the self-evolving network architecture, in-depth exploration of the competition and collaborative relationships between multi-level processes is achieved, thereby enhancing the accuracy of resource competition and bottleneck prediction.

[0146] In one implementation of the embodiment of the present application, the specific process of iterating the improved self-evolving network based on the benefit feedback of the game theory optimization strategy is as follows: Figure 6 As shown, Figure 6 A flowchart of iterating the improved self-evolving network based on the revenue feedback of the game theory optimization strategy provided in an embodiment of the present application includes:

[0147] S601. Define a deviation measurement function and track the degree of deviation between the prediction error and the return distribution.

[0148] Node benefits obtained based on the game theory optimization strategy in the above embodiment and network prediction error Based on this, we define the deviation measure function , which is used to characterize the cumulative deviation between global returns and prediction accuracy.

[0149] make represents the time variable, Indicates the current moment, and represents the weight factor for the forecast error and return distribution, and The power exponent of error and gain is calculated by the following formula: , integrating the deviations in multiple time intervals:

[0150] ;

[0151] in Indicates time The forecast error data, Indicates time According to the overall mean return obtained from game theory, The larger the value, the more significant the cumulative effect of forecast deviation is and the relatively insufficient return signal is, and the network needs to be further corrected.

[0152] S602: When a significant cumulative deviation occurs, the activation function or connection parameters of the hidden layer are adjusted for gain or attenuation, so that the network can take into account both global prediction performance and precise focus on key nodes.

[0153] Regarding the activation function in the hidden layer and the connection between network layers, based on Real-time evaluation of the balance between global operational efficiency and local key link response requirements. represents the activation function of the lth layer, Indicates the Connection parameters to the v-node, represents the correction step size, Represents the deviation amplification factor. Gain or attenuation is implemented on the activation function or connection parameters of a specific layer through the following formula:

[0154] ;

[0155] The activation function After each iteration, the deviation metric To update, if The larger the exponential factor Weaken the activation amplitude or connection strength to suppress excessive attention to areas with excessive global benefit deviation, and maintain relatively high-precision focus on key nodes, thereby taking into account both overall prediction accuracy and local bottleneck detection capabilities.

[0156] By introducing the above-mentioned deviation measurement mechanism and dynamic correction strategy into the self-evolving network, the network structure can be quickly adjusted when monitoring the prediction error and the deviation of the return distribution, so that specific key links can be focused on with high precision while taking into account the global prediction performance.

[0157] In one implementation of the embodiment of the present application, the stable convergence fusion model is used to output the congestion risk value of each node at a future moment and the execution time impact range of the entire process, thereby locating the root cause of the multi-level process. The specific implementation process of this process is as follows: Figure 7 As shown, Figure 7 A flowchart for predicting the future path of a multi-level process using a fusion model provided in an embodiment of the present application includes:

[0158] S701: Determine whether the self-evolving network is stable.

[0159] Based on the modified self-evolving network in the above embodiment, the training error is monitored. and earnings signals The trend of change over several iterations is used to determine whether the overall network has stabilized in parameter space. This state is used as the final output of the fusion model, solidifying the hidden layer connection weights and activation function configurations to obtain a converged fusion model with dynamic game optimization capabilities.

[0160] S702: Utilize the converged fusion model to perform node prediction.

[0161] Based on the converged fusion model, the node set of the multi-level process In the future The congestion probability is measured and the timeout risk is quantitatively inferred.

[0162] make Representation node In time The load occupancy rate, Representation node Resource capacity, overall threshold Used to characterize saturation values, is the exponential amplification factor, is a smoothing correction term. By calculating the node At the moment Congestion probability , and further evaluate the impact of downstream queuing or timeouts on the overall process based on the relationship between congestion probability and process structure:

[0163] ;

[0164] In the above formula, Indicates the time interval The cumulative amount of internal load occupancy, Used to prevent distortion of congestion probability when the load is low, Represents an exponential operation, which is used to amplify or suppress the impact of resource abundance on node congestion.

[0165] The higher the node In the subsequent moments The greater the congestion risk, the more significant the congestion risk. After determining the congestion probability of each node, the directed graph information of the process structure is combined with the execution order between nodes to infer the risk range of downstream links being affected by transmission queues or delays. Based on the time series, the impact on the overall process execution time is inferred.

[0166] S703. Output the result.

[0167] Combining the congestion probability and delay transmission analysis results in the above steps, a panoramic prediction output is generated, including the congestion probability value of each process link at future moments, the expected impact range on the overall execution time, and a ranked list of key risk nodes.

[0168] The above embodiment introduces a multi-level process node prediction method in this application. In order to make the implementation process of the solution in this application easier to understand, the above method in this application is introduced below in combination with specific application scenarios:

[0169] In this example, a large financial services group planned to deploy a comprehensive, multi-level process management and control system for its customer withdrawal process. In recent years, with the surge in customer withdrawal demand, each withdrawal process has repeatedly encountered resource competition and collaboration bottlenecks between parent, child, and grandchild processes in its daily operations. To ensure that the customer withdrawal process can continue to operate smoothly and efficiently under high concurrency conditions, the group decided to adopt the multi-level process node prediction method proposed in this invention to upgrade the entire process, including equipment data upload, lease audit, manual review, contract signing, and enterprise four-factor authentication.

[0170] First, a configuration application for the three-level customer withdrawal process is submitted through the online management platform of the present invention. The system first verifies the submitter's account and role permissions and structures the uploaded key indicator data (including the number of device data uploaded, lease review time, number of manual reviews, contract service load, and the company's four-factor authentication status) into a multidimensional monitoring sequence. Next, the system uses a formula combining cumulative summation and exponential decay to calculate the weighted monitoring value of each process, automatically identifying and marking abnormal nodes that deviate significantly from the normal range. Furthermore, the system performs deduplication processing on duplicate monitoring records under the same time index to prevent redundant data interference. After data cleaning is completed, the system normalizes the monitoring matrix, including the parent process, child processes, and grandchild processes, to form the initial data set required by the present invention.

[0171] After obtaining the initial data set, the system constructed adjacency matrices for the parent process, child process, and grandchild process based on the process execution order and resource dependencies. It also introduced a cross-layer coupling matrix and layer weight factors to achieve a global representation of the three-level process. Initial training of the self-evolving network was initiated. In the hidden layer, this network uses a piecewise cost function, comprehensively considering the balance between prediction error and network scale. When the prediction accuracy of a process node is insufficient and the network complexity remains controllable, the self-evolution algorithm automatically adds hidden layers or parameter units to strengthen the fitting ability of abnormal nodes. When withdrawal demand is relatively stable and node prediction errors are at a low level, the network shrinks its structure to reduce operating costs. Because dynamic tuning is achieved over multiple time steps, the system can quickly identify potential bottlenecks and promptly adjust corresponding child processes or grandchild processes even when faced with sudden withdrawal peaks.

[0172] The system then further analyzes the degree of resource competition among sub-processes and grand-child processes within each withdrawal process using a game theory optimization strategy. For nodes with high resource utilization but significant impact on subsequent processes, the system automatically amplifies their revenue signals, directing the network's hidden layer to focus on the status of these high-load nodes and providing early warnings of potential execution bottlenecks. For example, during the device data upload process, if a large number of uploads occur in a short period of time and server storage space is insufficient, the platform will mark it as a "high-risk red node" and automatically increase storage capacity to prevent insufficient space. Another example is that during the rental property review process, if an external system (the rental property review system) fails, causing the rental property identification function to fail, the predictive method can directly identify the cause of the error and request that the third-party system restore service as soon as possible. Furthermore, if the automatic review fails, the process will be transferred to manual review, which is less efficient. By monitoring the number of reviews, the predictive method can predict the need for additional reviewers and make advance staffing arrangements.

[0173] During the contract signing process, the program generates PDF files that require stamping and signatures, placing significant pressure on the system. By combining prediction methods with server load monitoring, if a large number of contract signing requests are received within a short period of time, R&D personnel are alerted to expand the contract signing service capacity to prevent customers from being stuck and unable to sign contracts. Finally, during the enterprise four-factor authentication process, if insufficient funds are encountered in the external system eSignPro, causing authentication failure, the full-path prediction method detected that many processes were stuck at this point, predicting that the company may be running out of funds.

[0174] After training and iteration, the self-evolving network converged. During peak withdrawal periods, the platform quantified the congestion probability and timeout risk of each withdrawal process, providing a key basis for allocating resources for the next day. If a node is diagnosed as potentially overloaded during peak hours, the system will temporarily add child or grandchild process instances the previous night. Leveraging resource isolation, external calls can be reliably completed, avoiding global congestion caused by accumulated error retries. Based on the final prediction and assessment, the platform also generates a comprehensive risk analysis report, ranking the key nodes within parent, child, and grandchild processes most likely to trigger chain reactions, allowing managers to conduct targeted pre-emptive resource scheduling and operational optimization.

[0175] To evaluate the effectiveness of the present invention's approach, the Group compared the traditional model with the present invention on several key indicators. The main results are as follows:

[0176] Monitoring accuracy:

[0177] Traditional methods: Most of them use offline statistics, and can only obtain execution results after the fact, which easily misses the opportunity for real-time warning.

[0178] The method of the present invention utilizes cumulative summation combined with exponentially decaying weighted monitoring values ​​to improve the accuracy of capturing abnormal nodes by about 40%.

[0179] Dynamic expansion efficiency:

[0180] Traditional method: When withdrawal demand fluctuates, new sub-processes need to be manually added or system configurations need to be changed, with an average response cycle of more than 2 days.

[0181] The method of the present invention: after detecting a high load, the self-evolving network can automatically complete the hidden layer expansion and sub-process instantiation within a few hours to achieve rapid adaptation.

[0182] Resource allocation and error retry:

[0183] Traditional method: Abnormal nodes are often difficult to capture in a timely manner and require manual investigation and re-pushing one by one.

[0184] The method of the present invention: the hidden layer correction driven by game theory benefits can accurately locate high-occupancy nodes and automatically inject retry services, saving more than 60% of operation and maintenance time.

[0185] Customer Satisfaction:

[0186] Traditional methods: Withdrawal delays and process bottlenecks occur frequently, and the average customer satisfaction rating is 3.6 / 5.

[0187] The method of the present invention: automatic expansion and flexible scheduling during peak periods ensure the continuity of the withdrawal process, and customer satisfaction is improved to 4.5 / 5.

[0188] Overall, compared with traditional decentralized process management methods, the solution of the present invention significantly improves the real-time response and bottleneck identification capabilities in complex withdrawal scenarios, shortens the time for manual intervention and repeated troubleshooting, and can still stably and accurately complete the customer withdrawal process in a high-concurrency environment.

[0189] It can be seen that the present invention has extremely broad application prospects and commercial value in multiple business scenarios, especially in the management field involving high coupling of parent processes, child processes and grandchild processes.

[0190] The above are some specific implementations of a multi-level process node prediction method provided by the embodiment of the present application. Based on this, the present application also provides a corresponding device. The device provided by the embodiment of the present application will be introduced from the perspective of functional modularization.

[0191] Figure 8 A schematic diagram of the structure of a multi-level process node prediction device provided in an embodiment of the present application. Figure 8 As shown, the multi-level process node prediction device 800 provided in the embodiment of the present application includes:

[0192] An acquisition unit 810 is configured to acquire the operation data of each level of the target level process, and obtain an initial data set based on the operation data of each level of the process;

[0193] A first constructing unit 820 is configured to construct a multi-level process comprehensive matrix based on the initial data set, wherein the multi-level process comprehensive matrix is ​​configured to represent the relationship between the processes at each level;

[0194] A second construction unit 830 is configured to construct an improved self-evolving network architecture based on the multi-level process comprehensive matrix; and to incorporate a game theory optimization strategy into the hidden layer of the improved self-evolving network;

[0195] An iterative optimization unit 840 is configured to iterate the improved self-evolving network based on the benefit feedback of the game theory optimization strategy, and obtain a stable convergent fusion model after multiple rounds of iterations;

[0196] The prediction unit 850 is configured to perform a panoramic prediction of the future path of the multi-level process using the stable convergence fusion model.

[0197] In one implementation of the embodiment of the present application, the acquisition unit is specifically configured to collect four monitoring indicators, namely, execution time, waiting queue length, external system call delay, and number of retries, for the parent process, child process, and grandchild process, respectively;

[0198] The monitoring indicators are recorded as a multidimensional sequence in chronological order, and weight factors of the parent process, child process and grandchild process are introduced to calculate the weighted monitoring value by combining cumulative summation with exponential decay;

[0199] The threshold strategy is used to mark abnormal points and remove duplicates of the monitoring values ​​that deviate from the mean, and then the obtained monitoring data matrix is ​​normalized to obtain an initial data set with a unified value range.

[0200] In one implementation of the embodiment of the present application, the first construction unit is specifically configured to:

[0201] Selecting multiple process links of the parent process, child process and grandchild process as a node set;

[0202] Combining the order and resource dependency in the node set, an adjacency matrix is ​​defined for each layer of the process. Then, a cross-layer coupling matrix and a layer weight factor are introduced to perform weighted fusion on the adjacency relationships among the parent process, child process, and grandchild process to obtain a multi-level process comprehensive matrix.

[0203] Preset key indicators are added to each node in the multi-level process comprehensive matrix to form a node feature vector.

[0204] In one implementation of the embodiment of the present application, the second construction unit is configured to construct an improved self-evolving network architecture based on the multi-level process integration matrix, including:

[0205] In a preset nested network template, receiving input including node feature vectors and comprehensive matrices;

[0206] The prediction error and network structure complexity are measured using a piecewise defined adaptive network cost function;

[0207] When the cost function exceeds the set threshold, the network depth or parameter scale is dynamically expanded to improve the fitting ability. When the cost function is at a low level, the network depth is reduced accordingly to improve the operating efficiency.

[0208] In one implementation of the embodiment of the present application, the second construction unit incorporates a game theory optimization strategy into the hidden layer of the improved self-evolving network, including:

[0209] Treat each node in the parent process, child process and grandchild process as the game subject participating in resource competition or collaboration;

[0210] Construct a profit function for each node based on resource occupancy and service priority information;

[0211] The profit signal obtained by the profit function is used as an additional input to correct the hidden layer output, so as to enhance the network's ability to focus on and identify potential bottleneck nodes.

[0212] In one implementation of the embodiment of the present application, the iterative optimization unit is specifically configured to:

[0213] Define a deviation metric function and track the degree of deviation between the forecast error and the return distribution;

[0214] When the cumulative deviation is significant, the activation function or connection parameters of the hidden layer are adjusted by gain or attenuation, so that the network can take into account both global prediction performance and precise focus on key nodes.

[0215] In one implementation of the embodiment of the present application, the stable convergence fusion model is used to output the congestion risk value of each node at a future moment and the execution time impact range of the overall process, thereby locating the root causes that may arise in multi-level processes.

[0216] The embodiments of the present application also provide corresponding devices and computer storage media for implementing the solutions provided by the embodiments of the present application.

[0217] The device includes a memory and a processor, the memory is used to store instructions or codes, and the processor is used to execute the instructions or codes so that the device executes the method described in any embodiment of the present application.

[0218] The computer storage medium stores code, and when the code is executed, the device executing the code implements the method described in any embodiment of the present application.

[0219] Through the description of the above embodiments, it can be seen that those skilled in the art can clearly understand that all or part of the steps in the above-mentioned embodiment methods can be implemented by means of software plus a general hardware platform. Based on this understanding, the technical solution of the present application can be embodied in the form of a software product. The computer software product can be stored in a storage medium, such as a read-only memory (ROM) / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network communication device such as a router) to execute the methods described in various embodiments or certain parts of the embodiments of the present application.

[0220] It is understandable that in the specific implementation of this application, the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved, when the above embodiments of this application are applied to specific products or technologies, need to obtain user permission or consent, and the collection, use and processing of relevant data need to comply with the relevant laws, regulations and standards of relevant countries and regions.

[0221] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0222] It should also be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device and apparatus embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments. The device and apparatus embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components indicated as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0223] The above is merely one specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A multi-level process node prediction method, characterized in that: The method comprises: Obtaining operation data of each level process in the target level process, and obtaining an initial data set based on the operation data of each level process; Constructing a multi-level process comprehensive matrix based on the initial data set, wherein the multi-level process comprehensive matrix is ​​used to represent the relationship between the processes at each level; Constructing an improved self-evolving network architecture based on the multi-level process integration matrix; Incorporating a game theory optimization strategy into the hidden layer of the improved self-evolving network; The improved self-evolving network is iterated based on the benefit feedback of the game theory optimization strategy; a stably converged fusion model is obtained after multiple rounds of iteration, and the fusion model is used to perform a panoramic prediction of the future path of the multi-level process.

2. The method according to claim 1, characterized in that The step of obtaining the operation data of each level process in the target level process and obtaining the initial data set based on the operation data of each level process includes: For the parent process, child process, and grandchild process, four monitoring indicators are collected: execution time, waiting queue length, external system call delay, and number of retries; The monitoring indicators are recorded as a multidimensional sequence in chronological order, and weight factors of the parent process, child process and grandchild process are introduced to calculate the weighted monitoring value by combining cumulative summation with exponential decay; The threshold strategy is used to mark abnormal points and remove duplicates of the monitoring values ​​that deviate from the mean, and then the obtained monitoring data matrix is ​​normalized to obtain an initial data set with a unified value range.

3. The method according to claim 1, characterized in that The constructing of a multi-level process comprehensive matrix based on the initial data set includes: Selecting multiple process links of the parent process, child process and grandchild process as a node set; Combining the order and resource dependency in the node set, an adjacency matrix is ​​defined for each layer of the process. Then, a cross-layer coupling matrix and a layer weight factor are introduced to perform weighted fusion on the adjacency relationships among the parent process, child process, and grandchild process to obtain a multi-level process comprehensive matrix. Preset key indicators are added to each node in the multi-level process comprehensive matrix to form a node feature vector.

4. The method according to claim 1, wherein The improved self-evolving network architecture constructed based on the multi-level process comprehensive matrix includes: In a preset nested network template, receiving input including node feature vectors and comprehensive matrices; The prediction error and network structure complexity are measured using a piecewise defined adaptive network cost function; When the cost function exceeds the set threshold, the network depth or parameter scale is dynamically expanded to improve the fitting ability. When the cost function is at a low level, the network depth is reduced accordingly to improve the operating efficiency.

5. The method according to claim 1, wherein The incorporation of game theory optimization strategies into the hidden layer of the improved self-evolving network includes: Treat each node in the parent process, child process and grandchild process as the game subject participating in resource competition or collaboration; Construct a profit function for each node based on resource occupancy and service priority information; The profit signal obtained by the profit function is used as an additional input to correct the hidden layer output, so as to enhance the network's ability to focus on and identify potential bottleneck nodes.

6. The method according to claim 1, characterized in that The iterating of the improved self-evolving network based on the benefit feedback of the game theory optimization strategy includes: Define a deviation metric function and track the degree of deviation between the forecast error and the return distribution; When the cumulative deviation is significant, the activation function or connection parameters of the hidden layer are adjusted by gain or attenuation, so that the network can take into account both global prediction performance and precise focus on key nodes.

7. The method according to claim 1, characterized in that The stable convergence fusion model is used to output the congestion risk value of each node at a future moment and the impact range of the execution time of the entire process, thereby locating the root causes that may occur in multi-level processes.

8. A multi-level process node prediction device, characterized in that: The device comprises: an acquisition unit, configured to acquire operation data of each level of the target level process, and obtain an initial data set based on the operation data of each level of the process; A first construction unit is configured to construct a multi-level process comprehensive matrix based on the initial data set, wherein the multi-level process comprehensive matrix is ​​configured to represent the relationship between the processes at each level; A second construction unit is configured to construct an improved self-evolving network architecture based on the multi-level process comprehensive matrix; and to incorporate a game theory optimization strategy into the hidden layer of the improved self-evolving network; an iterative optimization unit, configured to iterate the improved self-evolving network based on the benefit feedback of the game theory optimization strategy, and obtain a stable convergent fusion model after multiple rounds of iterations; The prediction unit is used to use the stable convergence fusion model to perform a panoramic prediction of the future path of the multi-level process.

9. A computing device, characterized in that The computing device includes: a memory and a processor; The memory is used to store computer programs; The processor is configured to implement the method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images