Artificial intelligence-based business process intelligent management method and system

Abstract

The application discloses a business process intelligent management method and system based on artificial intelligence, and belongs to the technical field of artificial intelligence. The method comprises the following steps: in response to a starting instruction of a target business, collecting process instance data of the target business in a historical execution period and constructing a business execution portrait; extracting resource scheduling features and node transfer features in different business scenarios according to the business execution portrait; constructing a time sequence state transition matrix of the business process based on the resource scheduling features and the node transfer features; obtaining current real-time running parameters of the target business, inputting the current real-time running parameters into a pre-trained business process prediction model, and obtaining a predicted state sequence of the target business in a preset future time period; and calculating a resource demand change trend and generating a dynamic scheduling strategy according to the predicted state sequence and the time sequence state transition matrix. The application realizes accurate prediction and dynamic allocation of resources in a business process.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a business process intelligent management method and system based on artificial intelligence. Background Technology

[0002] Existing technologies for business process management often employ static resource allocation methods or scheduling rules based on simple threshold triggers. Once the target business is initiated, the system typically allocates resources according to a pre-set, fixed template, failing to perceive the dynamic evolution of the business process's internal state. This management model is prone to discrepancies between resource allocation and actual execution needs when facing complex and ever-changing business scenarios. This can lead to resource contention and bottlenecks at critical nodes during peak business periods, or significant resource idleness and low utilization during off-peak periods.

[0003] Existing process mining techniques can reconstruct process models from historical logs, but their analysis results often remain at the level of macro-level statistical characteristics of the process, such as average node execution time or resource usage frequency. These techniques fail to effectively combine the temporal persistence of resource occupancy within the process with the resource transfer relationships between nodes. When a process node releases resources, the total amount of that resource transferred to subsequent nodes lacks a quantifiable transfer path description, causing scheduling actions to lag behind the immediate changes in the process's state. Furthermore, existing predictive scheduling techniques generally extrapolate total resource amounts based on a single time series, failing to refine the resource demand distribution along different state transition paths within the business process. This results in predictive scheduling lacking resolution specific to the process execution path.

[0004] This application aims to address how to quantitatively describe the temporal transmission relationship of resources between nodes in a business process, and how to combine macroscopic predictions of process status with the resource transmission patterns between nodes to achieve resource demand trend analysis that is tied to the business process execution path. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent management method and system for business processes based on artificial intelligence. By constructing a state transition matrix that reflects the temporal transmission relationship of resources, and combining the matrix with the predicted state sequence of the business process, a dynamic scheduling strategy with path discrimination capability is generated.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides an intelligent business process management method and system based on artificial intelligence. By collecting process instance data of the target business within its historical execution cycle, a business execution profile capable of depicting the actual execution state of the business is constructed. Based on this profile, resource scheduling characteristics and node flow characteristics under different business scenarios are extracted, thereby constructing a temporal state transition matrix of the business process. The system can obtain the current operating parameters of the target business in real time and, using a pre-trained business process prediction model, predict the predicted state sequence of the target business within a preset future time period.

[0007] In some preferred implementations, to construct an accurate business execution profile, multiple process nodes within historical execution cycles are extracted from process instance data. The node identifier, trigger timestamp, and completion timestamp of each process node are obtained. The average and variance of the node dwell time for each process node are calculated and embedded as node attributes into an initial process topology graph constructed based on the dependencies between nodes, thus obtaining a process topology graph with temporal attributes. Combining this with the resource consumption vector obtained by classifying and summarizing each process node by resource type within the historical execution cycle, and associating this vector with the corresponding node in the aforementioned topology graph, a comprehensive business execution profile reflecting the business execution status is finally generated. This profile integrates multi-dimensional information such as process structure, time consumption, and resource requirements, providing a solid data foundation for subsequent analysis.

[0008] In a specific technical solution, when constructing a temporal state transition matrix based on resource scheduling characteristics and node flow characteristics, the resource allocation ratio of each process node under different business scenarios is extracted from the resource scheduling characteristics to determine the type and quantity of resources released after execution. Simultaneously, the average execution time of each process node is extracted from the node flow characteristics, and the execution time window of the node is determined by combining the dependencies between nodes. By calculating the amount of resources transferred from each process node to its successor, the state transition weight between any two process nodes with dependencies can be determined. All state transition weights are combined to form the temporal state transition matrix. This matrix accurately describes the dynamic patterns of resource transfer in the business process over time and between nodes.

[0009] As a technical solution of this invention, when using a pre-trained business process prediction model for prediction, the obtained real-time running parameters include the node identifier of the currently started process node, the execution duration, and the consumed amount of each resource. This business process prediction model is built based on a Long Short-Term Memory (LSTM) network, and its input layer dimension matches the number of parameter types in the running parameters. Internally, the model can predict the estimated remaining execution duration of each started process node based on the current node identifier and the execution duration, thereby predicting the amount of resources to be consumed at each time step within a preset future time period. Arranging these data in chronological order forms a predicted state sequence. This solution leverages the advantages of LTM networks in processing time-series data, making the prediction of future resource consumption more accurate.

[0010] Furthermore, when calculating resource demand trends and generating dynamic scheduling strategies based on the predicted state sequence and time-series state transition matrix, the quantity of resources to be consumed at each time step in the predicted state sequence is first accumulated by resource type to obtain the total resource demand at each time step. Based on the state transition weights in the time-series state transition matrix, the total resource demand at each time step is mapped to each state transition path to obtain the path resource demand, which is then merged by resource type to obtain the combined resource demand at each time step. By performing a differential operation on the combined resource demand at all time steps, the resource demand trend is obtained. Based on this trend, resource replenishment instructions are generated during periods of rising resource demand, and resource release instructions are generated during periods of declining resource demand, ultimately combined into a dynamic scheduling strategy. This process achieves a closed loop from prediction to decision-making, enabling resource scheduling to respond proactively to business changes.

[0011] In the detailed scheme for generating dynamic scheduling strategies, the system sets positive and negative thresholds for the rate of change of resource demand and calculates the rate of change of resource demand between adjacent time steps. Time steps with a rate of change greater than the positive threshold are marked as rising indicators. During consecutive periods marked as rising indicators, the resource gap is calculated based on the combined resource demand and the current total available resources for each time step, thereby generating resource replenishment instructions containing resource type and quantity. Correspondingly, time steps with a rate of change less than the negative threshold are marked as falling indicators. During consecutive periods marked as falling indicators, the initial surplus of the total available resources and demand is calculated and compared with the resource occupancy limit of the process node to obtain the resource surplus, thereby generating resource release instructions containing information on resources to be released. The replenishment and release instructions with time information are arranged in chronological order to form a complete dynamic scheduling strategy with early warning and adjustment functions, effectively avoiding the risk of resource idleness or depletion.

[0012] As a preferred approach, the training process for the aforementioned business process prediction model involves collecting historical execution records of the target business across multiple historical execution cycles. These records contain the execution parameters and corresponding real states for each historical time step. The normalized historical execution parameters are used as training input samples, and the corresponding historical real states are used as training output labels, forming a dataset divided into training and validation sets. An initial model is constructed, comprising an input layer, an output layer, and a hidden layer composed of multiple long short-term memory units. The model's weights and bias parameters are iteratively updated on the training set using the backpropagation algorithm until the prediction error on the validation set meets the requirements, thus obtaining a pre-trained business process prediction model. This training method ensures that the model can effectively learn the complex dependencies and dynamic changes in the business process execution.

[0013] When extracting resource scheduling and node flow characteristics, the system reads the process topology diagram with time-series attributes to obtain the attributes and resource consumption vector of each node. Following the dependencies in the topology diagram, it traverses from the starting node to the ending node to determine the level number of each node. Based on the level number, it calculates the total resource consumption of all nodes within the same level and the change in resource consumption between different levels, combining these to obtain the resource scheduling characteristics. Simultaneously, it calculates the difference in average dwell time between each node and its successor nodes, arranging all differences into a sequence according to dependencies to form the node flow characteristics. The characteristics extracted in this way intuitively reflect the acceleration or delay effects of resource consumption patterns and execution rhythms at different stages of the business.

[0014] This invention also provides an intelligent business process management system based on artificial intelligence. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the aforementioned intelligent business process management method based on artificial intelligence. This system can achieve intelligent monitoring, accurate prediction, and adaptive resource scheduling of business processes, significantly improving the operational efficiency and resource utilization of business processes, and reducing process blockages and delays caused by resource mismatch.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows: By extracting the resource release type, resource release quantity, and node execution time window for each process node from resource scheduling characteristics and node flow characteristics, the resource transfer amount passed to subsequent nodes after each node's execution can be calculated. Based on this resource transfer amount, a state transition weight is constructed between any two process nodes with sequential dependencies, and all state transition weights are combined into a temporal state transition matrix. This matrix explicitly defines the quantitative flow rules of resources on the directed edges of the process topology, overcoming the limitation of traditional scheduling that only focuses on the total amount of the macro-level resource pool while ignoring the resource transfer paths between nodes. This allows the generation of scheduling strategies to accurately match the resource flow logic within the process.

[0016] After obtaining the predicted state sequence output by the business process prediction model, the amount of resources to be consumed at each time step in the sequence is mapped according to the weight relationship of each state transition path in the time-series state transition matrix to obtain the path resource demand for each state transition path. The demand for each path is merged and differentially calculated to depict the trend of resource demand changes associated with specific process execution paths. This trend not only reflects the fluctuations in the total future resources but also reveals which specific process state transition intervals the demand changes occur in. Based on this, resource replenishment instructions are generated during periods of increasing resource demand, and resource release instructions are generated during periods of decreasing resource demand. This allows the dynamic scheduling strategy to precisely intervene in the actual bottleneck transition paths within the process. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 It is a flowchart of an AI-based intelligent business process management method; Figure 2 This is a flowchart of the business execution profile construction and feature extraction process; Figure 3 This is a flowchart of the construction process for the temporal state transition matrix; Figure 4 This is a flowchart of a resource dynamic scheduling method based on business process prediction; Figure 5 This is a schematic diagram of the structure of an AI-based intelligent business process management system. Figure 6 It is a distribution chart of the dwell time of process nodes. Detailed Implementation

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

[0020] See Figure 1 This invention provides an intelligent business process management method based on artificial intelligence. The method includes: responding to the start command of a target business, collecting process instance data of the target business within its historical execution cycle, and constructing a business execution profile based on the process instance data; extracting resource scheduling features and node flow features of the target business under different business scenarios based on the business execution profile; constructing a temporal state transition matrix of the business process based on the resource scheduling features and node flow features; obtaining the current real-time operating parameters of the target business, inputting the current real-time operating parameters into a pre-trained business process prediction model to obtain a predicted state sequence of the target business within a preset future time period; calculating the resource demand change trend of the target business based on the predicted state sequence and the temporal state transition matrix, and generating a dynamic scheduling strategy based on the resource demand change trend.

[0021] Example 1: In specific implementation, refer to Figure 2 In response to the start command of the target business, the system collects process instance data of the target business within its historical execution cycle. The historical execution cycle refers to the multiple complete business process executions that the target business has completed before the start time. The process instance data contains detailed records of each execution of the target business within the historical execution cycle. Multiple process nodes of the target business within the historical execution cycle are extracted from the process instance data, and the node identifier, trigger timestamp, and completion timestamp of each process node are obtained. A process node is an indivisible execution step in the target business process. The node identifier uniquely identifies a process node, the trigger timestamp is the start time of the process node's execution, and the completion timestamp is the end time of the process node's execution.

[0022] After obtaining the node trigger timestamp and node completion timestamp for each process node, the node dwell time for each process node is calculated as follows: For a single process node, the time difference between its completion timestamp and trigger timestamp is taken as the node dwell time. For multiple node dwell times generated from multiple executions of the same process node within a historical execution cycle, the average and variance of the node dwell times are calculated. The average node dwell time is calculated using the following formula: ; in, Indicates the first The average node dwell time of each process node; Indicates the first The number of times each process node is executed within the historical execution cycle; Indicates the first The process node at the _ The node completion timestamp at the time of execution; Indicates the first The process node at the _ The node trigger timestamp at the time of execution. The variance of node dwell time is obtained by calculating the mean of the sum of squares of the differences between the node dwell time and the average value for each execution.

[0023] Based on the node identifier information of each process node and the execution order of process nodes recorded in the process instance data, the pre- and post-process dependencies between multiple process nodes are determined. Pre- and post-process dependencies indicate that a subsequent process node can only begin execution after the previous process node has completed. Based on these dependencies, an initial process topology graph for the target business is constructed. The initial process topology graph is a directed acyclic graph (DAG), where each vertex corresponds to a process node, and directed edges point from the direct predecessor process node to the direct successor process node, with the direction representing the pre- and post-process dependencies. During the actual construction process, all process nodes are traversed. For each process node, its node identifier information is added to the graph as a vertex, and directed edges are added based on the extracted pre- and post-process dependencies.

[0024] The average and variance of the node dwell time for each process node are used as node attributes for that node. These node attributes are then embedded into the corresponding node positions in the initial process topology graph, resulting in a process topology graph with temporal attributes. In some embodiments, the embedding method involves appending the average and variance of the node dwell time to the corresponding vertex data structure of the initial process topology graph, so that each vertex contains not only node identification information but also node attributes representing time consumption characteristics. The process topology graph with temporal attributes simultaneously expresses the structural relationships and time statistical characteristics of the target business.

[0025] The resource consumption records for each process node within its historical execution cycle are obtained. These records document the amount of various resources used or consumed by each process node during its execution. The resource consumption records are then categorized and summarized according to preset resource types to obtain a resource consumption vector for each process node. These preset resource types can be resource categories pre-defined based on the target business operating environment, such as computing resources, memory resources, storage read / write throughput resources, and network bandwidth resources. During the categorization and summarization, for each resource type, the consumption of that resource type by the process node across all executions within its historical execution cycle is summed or averaged, forming one dimension of the resource consumption vector. The number of dimensions in the resource consumption vector equals the number of preset resource types, and each component of a dimension corresponds to the historical average or cumulative consumption of a resource type.

[0026] The resource consumption vector is associated with the corresponding node in the process topology diagram with temporal attributes to generate a business execution profile. The association operation refers to combining the resource consumption vector of each process node as an additional attribute of the corresponding vertex with the vertex in the process topology diagram with temporal attributes, so that the business execution profile contains a complete description that integrates process structure, temporal statistical attributes, and resource consumption information.

[0027] After generating the business execution profile, the process topology diagram with time-series attributes is read from the business execution profile, and the node attributes and resource consumption vector of each process node in the process topology diagram with time-series attributes are obtained. The node attributes include the average and variance of the node dwell time, and the resource consumption vector includes the multi-dimensional resource consumption after classification and summarization.

[0028] Based on the dependencies in the process topology graph with temporal attributes, the process is traversed from the starting process node to the ending process node. During the traversal, the hierarchical position of each process node in the traversal path is recorded, and the node hierarchical number of each process node is determined according to its hierarchical position. The starting process node is a process node without any predecessor process node, and the ending process node is a process node without any successor process node. The hierarchical position is determined as follows: the node hierarchical number of the starting process node is set to 0; for any non-starting process node, its node hierarchical number is the maximum value of the node hierarchical numbers of all direct predecessor process nodes plus 1. A single topology traversal can assign node hierarchical numbers to all process nodes.

[0029] Based on the node hierarchy number and resource consumption vector of each process node, the total resource consumption of all process nodes within the same hierarchy and the change in resource consumption between different hierarchical levels are calculated. The total resource consumption of all process nodes within the same hierarchy is obtained by adding the resource consumption vectors of each process node with the same node hierarchy number according to the resource type dimension, resulting in a multi-dimensional sum vector corresponding to that hierarchy. The change in resource consumption between different hierarchical levels is obtained by subtracting the total resource consumption vector of the previous level from the sum vector of the next level, resulting in a multi-dimensional difference vector. The total resource consumption and the change in resource consumption are combined to form a resource scheduling feature. In some embodiments, the resource scheduling feature includes a sequence of total resource consumption arranged in hierarchical order and a sequence of change in resource consumption arranged in inter-hierarchical order.

[0030] Based on the node attributes and dependencies of each process node, calculate the time difference between the average node dwell time of each process node and the average node dwell time of its successor process nodes. For any directed edge from the i-th process node to the j-th process node, extract the average node dwell time of the i-th process node. The average node dwell time of the j-th process node ,calculate The duration differences are represented as the duration differences on the corresponding dependent edges. All duration differences are arranged according to their dependencies to form a duration difference sequence. This sequence corresponds one-to-one with the directed edges, and the positions of the elements in the sequence match the order of the dependent edges in the process topology graph. This duration difference sequence is used as a node flow feature. The node flow feature reflects the propagation of execution duration changes between process nodes.

[0031] Through the above process, the AI-based intelligent business process management method constructs a business execution profile containing structural and resource consumption information from historical process instance data, and further extracts resource scheduling features and node flow features for subsequent construction of time-series state transition matrices and predictive scheduling.

[0032] See Figure 6 The figure shows the distribution of node dwell time for each process node in the target business process in Example 1 using a scatter plot. The horizontal axis represents the process nodes, from process node 01 to process node 06; the vertical axis represents the node dwell time in seconds. Different shapes of scatter points correspond to different process nodes, and the legend clearly indicates the scatter point symbol corresponding to each node.

[0033] As shown in the graph, the dwell time of process node 01 is mainly concentrated between approximately 2.0 seconds and 2.7 seconds, with a relatively concentrated distribution and small fluctuations. The dwell time of process node 02 is concentrated between approximately 3.0 seconds and 5.2 seconds, with an average duration significantly longer than node 01, and the data distribution is slightly more dispersed, reflecting that the execution time of node 02 is relatively longer and more volatile. The dwell time of process node 03 ranges from approximately 1.3 seconds to 2.5 seconds, which is generally low and concentrated, indicating that this node executes quickly and stably.

[0034] The dwell time of process node 04 has the widest range, approximately 3.2 seconds to 7.4 seconds, with many high-duration values, indicating that node 04 has the longest execution time and significant fluctuations, potentially representing a critical bottleneck node in the business process. The dwell time of process node 05 is concentrated between approximately 3.8 seconds and 4.8 seconds, with a moderate and relatively even duration, indicating a moderate execution time. The dwell time of process node 06 is mainly distributed between approximately 1.5 seconds and 2.6 seconds, concentrated with minimal fluctuations, indicating a shorter execution time.

[0035] Example 2: In specific implementation, refer to Figure 3 The resource allocation ratio for each process node in the target business under different business scenarios is extracted from the resource scheduling features. The resource scheduling features include the resource consumption vector of each process node and the total resource consumption vector of all process nodes within the same level. For each process node, the resource consumption vector corresponding to that process node in the resource scheduling features is obtained, and the total resource consumption vector of the level to which that process node belongs is also obtained. The level division is based on the node level number determined when traversing from the initial process node. The resource allocation ratio is calculated one by one according to the preset resource types: for a resource type, the component of the process node's resource consumption vector corresponding to that resource type is divided by the component of the total resource consumption vector of the level corresponding to that resource type; the resulting ratio is the resource allocation ratio of the process node for that resource type. If the component of a certain resource type in the total resource consumption vector of the level is zero, then the resource allocation ratio for that resource type is 0.

[0036] Based on the resource allocation ratio, determine the type and quantity of resources released after each process node completes execution. Obtain the total available system resources for the target service. The total available system resources represent the upper limit of resources that the target service can allocate in the operating environment, and are recorded according to preset resource types. The total available system resources can be obtained from the global resource capacity information collected during business execution profile generation, or set by the system administrator based on the actual hardware configuration. For each resource type, multiply the total available system resources of the corresponding resource type by the resource allocation ratio of the process node for that resource type to obtain the resource occupancy quantity of the process node for that resource type. The process node occupies these resources during execution; upon completion, all occupied resources are released. Therefore, the released resource type is the resource type itself, and the released resource quantity is the resource occupancy quantity of the process node for that resource type. Perform the above operations on all preset resource types to obtain the set of released resource types and the corresponding set of released resource quantities for the process node.

[0037] The average execution time of each process node in the target business is extracted from the node flow characteristics. The node flow characteristics are stored as a sequence of duration differences during generation, and the average node dwell time of each process node is also stored in association. The average execution time value bound to each process node is read from the node flow characteristics; this value is equal to the average node dwell time of that process node within its historical execution cycles.

[0038] The execution time window for each process node is determined based on its average execution time and dependencies. Dependencies are recorded in the directed edges of the initial process topology graph, defining the predecessor-successor order between process nodes. The start time of the initial process node is set to 0, and its end time is the average execution time of the initial process node. For each process node other than the initial node, all its direct predecessor process nodes are identified, and their execution end times are obtained. The maximum value among these execution end times is taken as the start time of that process node. The end time of the process node is obtained by adding its average execution time to its start time. The interval formed by the start and end times constitutes the execution time window for that process node.

[0039] Based on the resource release type, resource release quantity, and execution time window of each process node, calculate the resource transfer amount that each process node passes to subsequent process nodes after execution. For any process node, obtain all its direct successor process nodes, which are determined by their dependencies. If the number of direct successor process nodes is 0, then this process node does not pass resources to its successors, and the resource transfer amount is 0. If the number of direct successor process nodes is greater than 0, then divide the resource release quantity of each type of process node by the number of direct successor process nodes to obtain the resource transfer quantity of that type obtained by each direct successor process node from the process node. Perform the above operation on all resource release types to form a resource transfer quantity vector from the process node to each direct successor process node. The dimension of the vector is consistent with the preset number of resource types, and each component of the dimension is the resource transfer quantity of the corresponding resource type.

[0040] State transition weights are constructed between any two process nodes with a sequential dependency relationship in the target business based on resource transfer volume. Pre-defined resource type weight coefficients are introduced, with each resource type corresponding to a weight coefficient ranging from 0 to 1, and the sum of all resource type weight coefficients being 1. These weight coefficients are pre-set by analyzing the impact of each resource type on process transition bottlenecks. For two process nodes with a sequential dependency relationship (i.e., a direct dependency from process node p to process node q), the resource transfer volume vector from process node p to process node q is obtained. The component corresponding to each resource type in the resource transfer volume vector is multiplied by the weight coefficient of that resource type and then summed to obtain the state transition weight from process node p to process node q. For two process nodes without a sequential dependency relationship, the state transition weight is set to 0. The state transition weights of all process node pairs are combined into a matrix, where the rows and columns of the matrix correspond to the process nodes in the target business. The element in the p-th row and q-th column of the matrix represents the state transition weight from process node p to process node q. This matrix is ​​the temporal state transition matrix. The time-series state transition matrix is ​​subsequently used to map the total resource demand to each state transition path to help calculate the trend of resource demand changes.

[0041] Example 3: In specific implementation, historical process operation records of the target business are collected over multiple historical execution cycles. Each historical process operation record contains historical operation parameters for a historical time step and the corresponding historical actual state. A historical time step refers to the time point at which the historical execution cycle is discretized and sampled. The time interval between adjacent historical time steps is a preset time step length, which is pre-set according to the process execution granularity of the target business. Historical operation parameters include the node identifier of the started process node in each historical time step, the execution duration of the started process node, and the amount of resources consumed. The historical actual state refers to the state sequence formed by arranging the actual amount of resources to be consumed in a preset future time period after the historical time step according to the time steps.

[0042] The historical operation parameters are normalized to obtain normalized historical operation parameters. The normalization process uses a maximum-minimum normalization method, linearly mapping each parameter in the historical operation parameters to the [0,1] interval. The normalized historical operation parameters are used as training input samples, and the corresponding historical true states are used as training output labels. The training input samples and training output labels form a training dataset, which is then divided into a training set and a validation set according to a preset ratio. The preset ratio is set as follows: the training set accounts for 80% of the total number of historical process operation records, and the validation set accounts for 20% of the total number of historical process operation records.

[0043] An initial business process prediction model is constructed, comprising an input layer, a hidden layer, and an output layer. The dimension of the input layer is consistent with the number of parameter types in the historical execution parameters. The parameter types in the historical execution parameters include: the number of bits in the node identifier of the started process node, the execution duration of the started process node, and the consumption value corresponding to each resource type. The sum of all these types constitutes the dimension of the input layer. The hidden layer contains multiple Long Short-Term Memory (LSTM) units, the number of which is preset based on the complexity of the target business process. Each LSM unit consists of a forget gate, an input gate, and an output gate. The forget gate controls the degree to which the unit state from the previous time step is forgotten, the input gate controls the degree to which current input information is added to the unit state, and the output gate controls the degree to which the unit state is output to the current hidden state. The dimension of the output layer is consistent with the dimension of the historical true state. The output layer is a fully connected layer, outputting the predicted value of the amount of resources to be consumed at each time step within a preset future time period.

[0044] During model training, the training set is input into the initial business process prediction model, and the weight and bias parameters of the initial business process prediction model are iteratively updated using the backpropagation algorithm. The mean squared error loss function is used, calculating the mean squared error between the predicted state sequence output by the output layer and the true state sequence in the training output labels. The Adam optimizer is used, with a preset initial learning rate that gradually decreases during iteration based on the training progress using a decay strategy. The iteration stopping condition is as follows: after each iteration, the validation set is input into the current model, the prediction error on the validation set is calculated, and iteration stops when the prediction error on the validation set is less than a preset error threshold, resulting in the pre-trained business process prediction model. The preset error threshold is set to a specific value based on the actual accuracy requirements of the target business.

[0045] After the pre-trained business process prediction model is trained, the current real-time operating parameters of the target business are obtained. These parameters include the node identifiers of the currently initiated process nodes, the execution duration of the currently initiated process nodes, and the current resource consumption. The node identifiers of the currently initiated process nodes are converted to numerical representation using the same encoding method as the training data; the execution duration of the currently initiated process nodes is the time difference from the node trigger timestamp to the current time; and the current resource consumption records the quantity of each type of resource consumed.

[0046] The current real-time operating parameters are input into the pre-trained business process prediction model. The input layer dimension of the business process prediction model is consistent with the number of parameter types in the current real-time operating parameters, and the number of parameter types in the current real-time operating parameters is the same as the number of parameter types during training. The input layer receives the normalized current real-time operating parameters and passes them to the hidden layer. The Long Short-Term Memory (LSTM) units in the hidden layer receive the current real-time operating parameters and, together with the previous LSM unit's internal state and the hidden state, calculate the current hidden state.

[0047] In the business process prediction model, the estimated remaining execution time of each started process node is predicted based on its node identifier and execution duration at the current moment. For each started process node, the feature representation corresponding to that process node is extracted from the model's hidden state representation. The output value is then calculated through a fully connected sub-network dedicated to remaining execution time prediction. This output value is processed by a non-negative activation function to obtain the estimated remaining execution time. This process can be represented by the following formula: ; in, Indicates the first The estimated remaining execution time for each started process node; This indicates the corresponding number in the hidden layer of the business process prediction model. Hidden state vectors generated by each started process node; This represents the weight matrix used to calculate the expected remaining execution time, which is updated through backpropagation during model training. This represents the bias vector used to calculate the expected remaining execution time, which is also learned during training; This indicates that the linear unit activation function is modified to ensure that the expected remaining execution time is non-negative.

[0048] Based on the estimated remaining execution time and the current resource consumption, the amount of resources to be consumed at each time step within a future preset time period is predicted for each started process node. The estimated remaining execution time of each started process node is combined with the current resource consumption to construct a time series input with a duration equal to the estimated remaining execution time. This time series input is then fed into the forward propagation path of the business process prediction model's output layer. The output layer outputs the amount of resources to be consumed at each time step within the future preset time period. The amount of resources to be consumed corresponds to all preset resource types, and the prediction output for each time step is a multi-dimensional vector, with the dimension equal to the number of preset resource types.

[0049] The quantities of resources to be consumed at each time step are arranged chronologically to obtain a predicted state sequence for the target service within a preset future time period. Each element in the predicted state sequence corresponds to a future time step and includes the quantity of resources to be consumed for each resource type at that time step. The predicted state sequence provides input data for subsequent calculations of resource demand change trends and the generation of dynamic scheduling strategies.

[0050] Example 4: In specific implementation, refer to Figure 4The process involves summing the resources to be consumed at each time step in the predicted state sequence according to resource type to obtain the total resource requirement for each time step. The predicted state sequence is the output of the business process prediction model and contains multiple consecutive future time steps. Each time step corresponds to a specific moment in time, and each time step records the expected quantity of various resources consumed by all started process nodes in the target business at that moment. Preset resource types include computing resources, memory resources, storage read / write throughput resources, and network bandwidth resources, each corresponding to a dimension. For the u-th time step in the predicted state sequence, the resources to be consumed by all process nodes at that time step are extracted. The quantities for each resource type are summed separately, and all quantities of the same resource type are added together to obtain a multi-dimensional vector. Each component of this vector represents the total resource requirement for the corresponding resource type at the u-th time step. This process is repeated for each time step within a preset future time period to obtain the total resource requirement vector for each time step.

[0051] Based on the state transition weights between process nodes in the time-series state transition matrix, the total resource requirement for each time step is mapped to the corresponding state transition path in the time-series state transition matrix, resulting in the path resource requirement for each time step on each state transition path. The time-series state transition matrix is ​​a square matrix, where rows and columns correspond to all process nodes in the target business. The element in the p-th row and q-th column represents the state transition weight from the p-th process node to the q-th process node. When there is a dependency relationship between the p-th and q-th process nodes, the state transition weight is a value greater than 0 and less than or equal to 1; when there is no dependency relationship, the state transition weight is 0. The mapping process is as follows: for the total resource requirement vector of a certain time step, traverse each non-zero state transition weight in the time-series state transition matrix, multiply the state transition weight by the total resource requirement vector of that time step, and obtain the path resource requirement vector corresponding to that state transition path. The path resource requirement vector also has the same dimension as the number of resource types. For process node pairs with a state transition weight of zero, no path resource requirement is generated.

[0052] For each time step, all path resource requirements are merged according to resource type to obtain the merged resource requirement for each time step. Within the same time step, the path resource requirement vectors generated on all state transition paths are added together dimension by dimension according to resource type. The components of all path resource requirement vectors in each resource type dimension are summed to obtain a merged multidimensional vector, which is the merged resource requirement for that time step. This merging operation is performed sequentially for each time step within a preset future time period to obtain the merged resource requirement vector for each time step.

[0053] The combined resource demand for all time steps within a preset future time period is differentially calculated sequentially to obtain the resource demand trend for the target business. There are U time steps within the preset future time period. Differential calculations are performed on the combined resource demand vectors for the u-th and u+1-th time steps. The difference is obtained by subtracting the combined resource demand vector of the previous time step from the vector of the next time step. Differential calculations are performed sequentially on u from 1 to U-1, resulting in U-1 difference vectors. This set of difference vectors, arranged in chronological order, constitutes the resource demand trend.

[0054] Set positive and negative thresholds for the resource demand change rate, and calculate the resource demand change rate between each adjacent time step in the resource demand change trend. The positive threshold is 0.15, and the negative threshold is -0.15. The positive and negative thresholds are set based on: obtaining historical resource demand data for the target business during normal and stable operation periods in multiple historical execution cycles, calculating the sequence of the total resource demand change rate between adjacent historical time steps, taking the 90th percentile value of this sequence as the positive threshold, and taking the 10th percentile value of this sequence as the negative threshold. The total resource demand change rate is calculated according to the following formula: ; in, This represents the rate of change in resource demand for the u-th adjacent time step, where u ranges from 1 to U-1; U is the total number of time steps within the preset future time period; and D represents the total number of preset resource types. This represents the component value of the merged resource demand at time step u+1 for the d-th resource type; This represents the component value of the merged resource demand at time step u for resource type d; This represents the total resource requirement scalar at time step u+1, which is the sum of all resource type components; This represents the total resource requirement scalar at the u-th time step; when the denominator... When the value is 0, the rate of change in resource demand for that adjacent time step is... The value is 0.

[0055] After calculating the rate of change of resource demand for each adjacent time step pair, time steps with a rate of change of resource demand greater than the positive threshold of 0.15 are marked as resource demand rising, time steps with a rate of change of resource demand less than the negative threshold of -0.15 are marked as resource demand falling, and time steps with a rate of change of resource demand between the positive threshold of 0.15 and the negative threshold of -0.15 and including the endpoints are marked as resource demand remaining stable.

[0056] Extracting consecutive time steps marked as increasing resource demand from the resource demand trend constitutes a period of increasing resource demand. A period of increasing resource demand consists of multiple consecutive time steps; in the resource demand trend sequence, adjacent time step pairs are all marked as increasing resource demand. During extraction, the marked sequence is scanned sequentially along the time steps, and the time step interval spanned by consecutively appearing increasing resource demand marks is considered a period of increasing resource demand. Similarly, extracting consecutive time steps marked as decreasing resource demand from the resource demand trend constitutes a period of decreasing resource demand.

[0057] During periods of rising resource demand, the total available resources at the starting time step of this period are obtained and used as the initial resource quantity. The total available resources refer to the total amount of each resource type that is not yet occupied in the target service's operating environment, provided by the system's real-time monitoring module. Starting from the starting time step of the rising resource demand period, the merged resource demand of each time step is traversed sequentially according to time. The merged resource demand of the current time step is compared with the remaining resources after the previous time step. The remaining resources after the previous time step are determined as follows: for the starting time step, the remaining resources are the same as the initial resource quantity; for non-starting time steps, the remaining resources are the remaining resources at the beginning of the previous time step minus the merged resource demand of the previous time step. When the merged resource demand of the current time step is greater than the remaining resources, the difference between the merged resource demand and the remaining resources is used as the resource gap for the current time step. The resource gap is a multi-dimensional vector, with each dimension corresponding to the gap quantity of a resource type. The resource gap amounts for all time steps within the period of increased resource demand are accumulated chronologically. This accumulation is done by adding the resource gap amounts for each time step separately for each resource type, resulting in the total resource gap for the period of increased demand. Based on the total resource gap, the required replenishment quantities for each resource type are determined, and a resource replenishment instruction is generated using the replenishment quantities for each resource type and the corresponding replenishment time step. The resource replenishment instruction includes the replenishment quantity for each resource type and the specific time step for executing the replenishment operation.

[0058] During periods of declining resource demand, obtain the total available resources at the starting time step of the period and use this total available resources as the initial available amount. Starting from the starting time step of the period of declining resource demand, iterate through the combined resource demand of each time step in chronological order. Using the initial available amount as the initial value, for each time step, calculate the difference between the total available resources and the combined resource demand of the current time step, and use this difference as the initial surplus for the current time step. The total available resources for the current time step are maintained as follows: for the starting time step, the total available resources for the current time step equals the initial available amount; for subsequent time steps, the total available resources for the current time step is the total available resources of the previous time step minus the combined resource demand of the previous time step, leaving the remaining resources. Obtain the resource occupancy limit for each process node corresponding to each time step during the period of declining resource demand. The resource occupancy limit for each process node is the maximum amount of resources allowed to be occupied during runtime, pre-set according to the business operating environment, and read from the business execution profile or system configuration parameters. The initial resource surplus at the current time step is compared with the resource utilization limit: if the initial surplus is greater than the resource utilization limit, the difference between the initial surplus and the resource utilization limit is taken as the resource surplus at the current time step; if the initial surplus is not greater than the resource utilization limit, the resource surplus at the current time step is set to zero. The resource surplus of all time steps within the resource demand decline period is accumulated chronologically, by adding the resource surplus of each time step separately according to resource type, to obtain the total resource surplus for the resource demand decline period. Based on the total resource surplus, the release quantity for each resource type is determined, and a resource release instruction is generated with the release quantity for each resource type and the corresponding release time step. The resource release instruction contains the release quantity for each resource type and the specific time step for executing the release operation.

[0059] The dynamic scheduling strategy is derived by arranging and combining resource replenishment and release instructions according to their respective time steps. The dynamic scheduling strategy is a list of time-series instructions, where each instruction explicitly indicates its execution time step, instruction type (replenishment or release), resource type, and number of operations. When both resource replenishment and release instructions exist at the same time step, the dynamic scheduling strategy includes both types of instructions for that time step, organized in the order of replenishment followed by release.

[0060] Example 5: See Figure 5An AI-based intelligent business process management system includes a memory, a processor, and a computer program stored in the memory and running on the processor. In specific implementations, the memory can be one or more of random access memory, read-only memory, and flash memory, and the processor can be one or more of a central processing unit, graphics processing unit, and tensor processor. When the processor executes the computer program, it implements the steps of an AI-based intelligent business process management method.

[0061] In practical implementation, the computer program is logically divided into a business execution profile construction unit, a feature extraction unit, a temporal state transition matrix construction unit, a business process prediction unit, and a dynamic scheduling strategy generation unit. The business execution profile construction unit responds to the target business's start command, collects process instance data of the target business within its historical execution cycle, and constructs a business execution profile based on this data. The unit extracts multiple process nodes of the target business within its historical execution cycle from the process instance data, obtaining the node identifier, trigger timestamp, and completion timestamp for each node. Based on the trigger and completion timestamps, the unit calculates the node dwell time for each process node and calculates the average and variance of the dwell time over the historical execution cycle. Finally, based on the node identifier information, the unit determines the dependencies between multiple process nodes and constructs an initial process topology diagram of the target business based on these dependencies. The business execution profile construction unit uses the average and variance of the node dwell time for each process node as the node attribute of the corresponding process node, and embeds the node attribute into the corresponding node position in the initial process topology diagram to obtain a process topology diagram with time-series attributes. The business execution profile construction unit obtains the resource consumption records of each process node within the historical execution cycle, classifies and summarizes the resource consumption records according to preset resource types, and obtains the resource consumption vector for each process node. The business execution profile construction unit associates the resource consumption vector with the corresponding node in the process topology diagram with time-series attributes to generate a business execution profile.

[0062] The feature extraction unit is used to extract resource scheduling features and node flow features of the target business under different business scenarios based on the business execution profile. The feature extraction unit reads the process topology diagram with temporal attributes from the business execution profile and obtains the node attributes and resource consumption vector of each process node in the process topology diagram with temporal attributes. Following the dependencies in the process topology diagram with temporal attributes, the feature extraction unit traverses from the starting process node to the ending process node, recording the hierarchical position of each process node in the traversal path during the traversal, and determining the node hierarchical number of each process node based on its hierarchical position. Based on the node hierarchical number and resource consumption vector of each process node, the feature extraction unit calculates the total resource consumption of all process nodes within the same hierarchical level and the change in resource consumption between different hierarchical levels, combining the total resource consumption and the change in resource consumption into resource scheduling features. The feature extraction unit calculates the time difference between the average node dwell time of each process node and the average node dwell time of subsequent process nodes based on the node attributes and dependencies of each process node. All time differences are arranged into a time difference sequence according to the dependencies, and the time difference sequence is used as the node flow feature.

[0063] The temporal state transition matrix construction unit is used to construct the temporal state transition matrix of a business process based on resource scheduling characteristics and node flow characteristics. First, it extracts the resource allocation ratio of each process node in the target business under different business scenarios from the resource scheduling characteristics, and determines the type and quantity of resources released by each process node after execution based on the resource allocation ratio. Second, it extracts the average execution time of each process node in the target business from the node flow characteristics, and determines the node execution time window for each process node based on the average execution time and dependencies. Third, it calculates the amount of resources transferred from each process node to subsequent process nodes after execution based on the resource release type, quantity, and execution time window. Finally, it constructs the state transition weights between any two process nodes with dependencies in the target business based on the resource transfer amount, and combines all state transition weights into a temporal state transition matrix.

[0064] The business process prediction unit acquires the current real-time operating parameters of the target business and inputs these parameters into a pre-trained business process prediction model to obtain the predicted state sequence of the target business within a preset future time period. The business process prediction unit loads the pre-trained business process prediction model, which is built based on a Long Short-Term Memory (LSTM) network. The pre-trained business process prediction model includes an input layer, hidden layers, and an output layer. The hidden layers contain multiple LSM units, each containing a forget gate, an input gate, and an output gate. The input layer dimension of the pre-trained business process prediction model corresponds to the number of parameter types in the current real-time operating parameters, and the output layer dimension corresponds to the number of time steps multiplied by the number of resource types within the preset future time period. The training process of the pre-trained business process prediction model is as follows: Historical process execution records of the target business are collected over multiple historical execution cycles. The historical execution parameters are normalized, and these normalized parameters are used as training input samples. The corresponding historical real-time states are used as training output labels, forming a training dataset divided into training and validation sets. An initial business process prediction model is constructed, and the weight and bias parameters are iteratively updated using the backpropagation algorithm until the prediction error on the validation set is less than a preset error threshold. The iteration then stops, resulting in the pre-trained business process prediction model. The current real-time execution parameters obtained by the business process prediction unit include the node identifier of the currently started process node, the execution duration of the currently started process node, and the amount of resources consumed at the current time. After the business process prediction unit inputs the current real-time execution parameters into the pre-trained business process prediction model, the model predicts the estimated remaining execution duration of each started process node based on the node identifier and execution duration, and predicts the amount of resources to be consumed by each started process node at each time step within a preset future time period based on the estimated remaining execution duration and the amount of resources consumed at the current time. The business process prediction unit arranges the amount of resources to be consumed at each time step in chronological order to obtain the predicted state sequence of the target business within a preset future time period.

[0065] The dynamic scheduling strategy generation unit calculates the resource demand change trend of the target service based on the predicted state sequence and the time-series state transition matrix, and generates a dynamic scheduling strategy based on the resource demand change trend. The dynamic scheduling strategy generation unit accumulates the amount of resources to be consumed at each time step in the predicted state sequence according to resource type to obtain the total resource demand at each time step. The accumulation process uses the following formula: ; in, Indicates the first In the first time step Total resource demand for each resource type; For time step index, The value ranges from 1 to positive integers, This represents the total number of time steps within a preset future time period. For resource type indexing, The value ranges from 1 to positive integers, The total number of preset resource types; Indicates the first The total number of process nodes that consume resources within a time step; Indicates the first The time step, the first The process node for the first The amount of resources to be consumed for each resource type Obtained from the predicted state sequence.

[0066] The dynamic scheduling strategy generation unit maps the total resource demand of each time step to the corresponding state transition path in the time-series state transition matrix according to the state transition weights between each process node, thus obtaining the path resource demand for each time step on each state transition path. The dynamic scheduling strategy generation unit merges all path resource demands for each time step according to resource type, obtaining the merged resource demand for each time step. The dynamic scheduling strategy generation unit performs differential operations on the merged resource demands of all time steps within a preset future time period in chronological order to obtain the resource demand change trend. The dynamic scheduling strategy generation unit sets positive and negative thresholds for the resource demand change rate, calculates the resource demand change rate between each adjacent time step in the resource demand change trend, marks time steps with a resource demand change rate greater than the positive threshold as resource demand increasing, time steps with a resource demand change rate less than the negative threshold as resource demand decreasing, and time steps with a resource demand change rate between the positive and negative thresholds as resource demand stable. The dynamic scheduling strategy generation unit extracts consecutive time steps marked as increasing resource demand from the trend of resource demand changes to form a resource demand increase period, and extracts consecutive time steps marked as decreasing resource demand to form a resource demand decrease period. Within a resource demand increase period, the dynamic scheduling strategy generation unit obtains the total available resources at the starting time step as the initial resource quantity. Starting from the starting time step, it sequentially traverses the combined resource demand of each time step, comparing the combined resource demand of the current time step with the remaining resources after the previous time step. If the combined resource demand of the current time step is greater than the remaining resources, the difference is taken as the resource gap for the current time step. These differences are accumulated to obtain the total resource gap. Based on the total resource gap, a resource replenishment instruction containing resource type, resource quantity, and a replenishment time step is generated. During periods of declining resource demand, the dynamic scheduling strategy generation unit obtains the total available resources at the starting time step of that period as the initial available amount. Starting from the starting time step, it sequentially iterates through the combined resource demand of each time step, calculating the difference between the total available resources and the combined resource demand as the initial surplus. It then obtains the resource occupancy limit for each process node at each time step. If the initial surplus exceeds the resource occupancy limit, the difference between the initial surplus and the resource occupancy limit is used as the resource surplus for the current time step. These are accumulated to obtain the total resource surplus. Based on the total resource surplus, a resource release instruction is generated, containing the type and quantity of resources to be released, as well as the release time step. The dynamic scheduling strategy generation unit then arranges and combines the resource replenishment and release instructions according to their respective time steps to obtain the dynamic scheduling strategy.

[0067] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A business process intelligent management method based on artificial intelligence, characterized in that, The method includes: In response to the start command of the target business, process instance data of the target business in the historical execution cycle is collected, and a business execution profile is constructed based on the process instance data; Based on the business execution profile, the resource scheduling features and node flow features of the target business under different business scenarios are extracted. Specifically, this includes: extracting multiple process nodes of the target business in the historical execution cycle from the process instance data, and obtaining the node identification information, node trigger timestamp and node completion timestamp of each process node. Based on the node trigger timestamp and node completion timestamp of each process node, calculate the node dwell time of each process node, and calculate the average and variance of the node dwell time of each process node in the historical execution cycle. The sequential dependencies between the multiple process nodes are determined based on the node identification information, and an initial process topology diagram of the target business is constructed based on the sequential dependencies. Based on the resource scheduling characteristics and the node flow characteristics, a temporal state transition matrix for the business process is constructed. Specifically, this involves: extracting the resource allocation ratio of each process node in the target business under different business scenarios from the resource scheduling characteristics, and determining the type and quantity of resources released by each process node after execution based on the resource allocation ratio; extracting the average execution time of each process node in the target business from the node flow characteristics, and determining the node execution time window of each process node based on the average execution time and the preceding and following dependencies; calculating the resource transfer amount that each process node passes to subsequent process nodes after execution based on the resource release type, resource release quantity, and node execution time window of each process node; constructing the state transition weight between any two process nodes with preceding and following dependencies in the target business based on the resource transfer amount, and combining all state transition weights into the temporal state transition matrix. The process involves obtaining the current real-time operating parameters of the target service and inputting these parameters into a pre-trained business process prediction model to obtain a predicted state sequence of the target service within a preset future time period. Specifically, this involves: obtaining the current real-time operating parameters of the target service, including the node identifier of the currently started process node, the execution duration of the currently started process node, and the amount of resources consumed at the current time; inputting these current real-time operating parameters into a pre-trained business process prediction model, which is constructed based on a Long Short-Term Memory (LSTM) network, and whose input layer dimension is consistent with the number of parameter types in the current real-time operating parameters; predicting the expected remaining execution duration of each started process node based on its node identifier and execution duration; predicting the amount of resources to be consumed at each time step of each started process node within a preset future time period based on its expected remaining execution duration and the amount of resources consumed at the current time; and arranging the amount of resources to be consumed at each time step in chronological order to obtain a predicted state sequence of the target service within a preset future time period. Based on the predicted state sequence and the time-series state transition matrix, the resource demand change trend of the target service is calculated, and a dynamic scheduling strategy is generated based on the resource demand change trend.

2. The intelligent business process management method based on artificial intelligence according to claim 1, characterized in that, The process of constructing a business execution profile based on the process instance data also includes: The average and variance of the node dwell time of each process node are used as the node attributes of the corresponding process node. The node attributes are embedded into the corresponding node positions of the initial process topology graph to obtain a process topology graph with time sequence attributes. Obtain the resource consumption record of each process node within the historical execution cycle, classify and summarize the resource consumption record according to the preset resource type, and obtain the resource consumption vector of each process node. The resource consumption vector is associated with the corresponding node in the process topology diagram with time sequence attributes to generate the business execution profile.

3. The intelligent business process management method based on artificial intelligence according to claim 2, characterized in that, The step of calculating the resource demand change trend of the target service based on the predicted state sequence and the time-series state transition matrix, and generating a dynamic scheduling strategy based on the resource demand change trend, specifically includes: The total resource requirement for each time step is obtained by summing up the resource requirements for each time step according to the resource type. According to the state transition weights between each process node in the time-series state transition matrix, the total resource requirement of each time step is mapped to the corresponding state transition path in the time-series state transition matrix to obtain the path resource requirement of each time step on each state transition path. The resource requirements of all paths at each time step are merged according to resource type to obtain the merged resource requirements for each time step. The combined resource requirements of all time steps within the preset future time period are differentially calculated in chronological order to obtain the resource requirement change trend of the target business. Based on the trend of resource demand changes, the periods of rising and falling resource demand are determined. During the periods of rising resource demand, resource replenishment instructions are generated, and during the periods of falling resource demand, resource release instructions are generated. The resource replenishment instructions and the resource release instructions are combined in chronological order to form the dynamic scheduling strategy.

4. The intelligent business process management method based on artificial intelligence according to claim 3, characterized in that, The process involves determining periods of rising and falling resource demand based on the trend of resource demand changes, generating resource replenishment instructions during periods of rising resource demand, generating resource release instructions during periods of falling resource demand, and combining the resource replenishment and release instructions in chronological order to form the dynamic scheduling strategy. Specifically: Set positive and negative thresholds for the rate of change of resource demand, and calculate the rate of change of resource demand between each adjacent time step in the trend of resource demand change. Time steps with a rate of change in resource demand greater than the positive threshold are marked as a sign of rising resource demand; time steps with a rate of change in resource demand less than the negative threshold are marked as a sign of falling resource demand; and time steps with a rate of change in resource demand between the positive and negative thresholds are marked as a sign of stable resource demand. Extract time steps continuously marked as indicators of rising resource demand from the trend of resource demand changes to form a period of rising resource demand; and extract time steps continuously marked as indicators of falling resource demand from the trend of resource demand changes to form a period of falling resource demand. During the period of increased resource demand, the resource gap is calculated based on the combined resource demand at each time step and the total available resources, and a resource replenishment instruction containing the resource type and quantity is generated based on the resource gap. During the period of decreased resource demand, the resource surplus is calculated based on the combined resource demand and the total amount of currently available resources at each time step during the period of decreased resource demand, and a resource release instruction containing the type and quantity of resources to be released is generated based on the resource surplus. The resource replenishment instructions and resource release instructions are arranged and combined according to the time sequence of their respective time steps to obtain the dynamic scheduling strategy.

5. The intelligent business process management method based on artificial intelligence according to claim 4, characterized in that, During the period of increased resource demand, a resource gap is calculated based on the combined resource demand at each time step and the total available resources. A resource replenishment instruction, including resource type and quantity, is then generated based on the resource gap. Specifically: Obtain the total amount of currently available resources at the starting time step within the period of rising resource demand, and use the total amount of currently available resources as the starting resource amount; Starting from the initial time step within the period of rising resource demand, the merged resource demand of each time step is traversed sequentially according to time order. The merged resource demand of the current time step is compared with the remaining resource quantity after the end of the previous time step. When the merged resource demand is greater than the remaining resource quantity, the difference between the merged resource demand and the remaining resource quantity is taken as the resource gap of the current time step. The resource gap amounts at all time steps during the period of increased resource demand are summed up in chronological order to obtain the total resource gap amount during the period of increased resource demand. The total resource shortage is used to determine the replenishment quantity of each resource type that needs to be replenished, and the resource replenishment instruction is generated based on the replenishment quantity of each resource type. The resource replenishment instruction includes the replenishment quantity of each resource type and the corresponding replenishment time step.

6. The intelligent business process management method based on artificial intelligence according to claim 5, characterized in that, During the period of decreased resource demand, a resource surplus is calculated based on the combined resource demand at each time step and the total available resources. A resource release instruction containing the type and quantity of resources to be released is then generated based on this resource surplus. Specifically: Obtain the total amount of currently available resources at the starting time step within the period of decreased resource demand, and use the total amount of currently available resources as the starting available amount; Starting from the initial time step within the period of decreased resource demand, the combined resource demand of each time step is traversed sequentially according to time order. The difference between the total available resources at the current time step and the combined resource demand at the current time step is calculated, and the difference is used as the initial surplus at the current time step. Obtain the upper limit of resource usage for each process node corresponding to each time step during the period of decreased resource demand. Compare the initial surplus of the current time step with the upper limit of resource usage. When the initial surplus is greater than the upper limit of resource usage, take the difference between the initial surplus and the upper limit of resource usage as the resource surplus of the current time step. The resource surplus of all time steps during the period of decreased resource demand is accumulated in chronological order to obtain the total resource surplus during the period of decreased resource demand. The release quantity of each resource type is determined based on the total resource surplus, and the resource release instruction is generated based on the release quantity of each resource type. The resource release instruction includes the release quantity of each resource type and the corresponding release time step.

7. The intelligent business process management method based on artificial intelligence according to claim 1, characterized in that, The training process of the pre-trained business process prediction model is as follows: Collect historical process operation records of the target business in multiple historical execution cycles. Each historical process operation record contains historical operation parameters of historical time steps and corresponding historical real status. The historical operating parameters are normalized to obtain normalized historical operating parameters, and the normalized historical operating parameters are used as training input samples, and the corresponding historical real states are used as training output labels. The training input samples and the training output labels are combined to form a training dataset, and the training dataset is divided into a training set and a validation set according to a preset ratio; An initial business process prediction model is constructed, which includes an input layer, a hidden layer, and an output layer. The hidden layer contains multiple long short-term memory units, and each long short-term memory unit contains a forget gate, an input gate, and an output gate. The training set is input into the initial business process prediction model, and the weight parameters and bias parameters of the initial business process prediction model are iteratively updated through the backpropagation algorithm until the prediction error of the initial business process prediction model on the validation set is less than a preset error threshold, and the iteration stops, thus obtaining the pre-trained business process prediction model.

8. An intelligent management system for business processes based on artificial intelligence, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent management method for business processes based on artificial intelligence as described in any one of claims 1 to 7.

Images