A real-time tracking method for the status of recharge orders

Abstract

The present invention discloses a method for real-time tracking of recharge order status, including: receiving standardized order data, constructing a state transition hypothesis set in combination with a pre-established order causal network model; performing non-linear backtracking analysis to generate a state backtracking result; performing a differential backtracking strategy to generate and execute an optimized backtracking plan; performing Bayesian network forward-backward inference fusion to generate a final state evaluation result. The present invention improves the accuracy and real-time performance of order status tracking in complex environments, especially performing excellently in the case of data delay and broken causal chains.

Description

Technical Field

[0001] The present invention belongs to the field of order status tracking and optimization, and in particular, to a method for real-time tracking of recharge order status. Background Art

[0002] In today's environment where digital payment is becoming increasingly popular, the real-time tracking of recharge order status has become a core function of electronic payment platforms. Accurately tracking the entire process status of an order from initiation to completion not only directly affects user experience and satisfaction but also relates to the operational efficiency and risk control capabilities of the platform. Especially in complex scenarios involving multiple payment channels and multi-system collaboration, real-time grasping of order status changes and rapid response to abnormal situations have become key indicators for the stability and reliability of payment systems.

[0003] Currently, the mainstream recharge order status tracking methods are mainly based on fixed state machine models and linear tracking mechanisms. These methods usually adopt predefined order status sets and conversion rules, use fixed time window aggregation to process multi-source data, and combine simple timeout retry mechanisms to handle data latency situations. Some advanced systems introduce machine learning techniques for status prediction, such as using recurrent neural networks to predict the possible next status of an order or adopting algorithms like isolation forest to detect abnormal status conversions. In the case of data inconsistency, a common method is the linear backtracking strategy based on timestamps, that is, updating the order historical status in reverse order according to the arrival time of data.

[0004] However, the existing technologies still have obvious deficiencies when facing complex payment environments. First, the fixed state space model cannot adapt to dynamically changing business scenarios, especially when dealing with abnormal states or intermediate states outside the preset state machine, and it performs poorly. Second, the traditional linear backtracking mechanism is difficult to handle complex topological structures with state forks and merges, and cannot accurately capture the non-linear dependence relationships between states, resulting in a significant increase in the state judgment error rate in cases of severe data latency or permanent loss of some data. Summary of the Invention

[0005] The object of the invention is to provide a method for real-time tracking of recharge order status, hoping to solve at least one technical problem existing in the prior art.

[0006] Technical solution: A method for real-time tracking of recharge order status includes the following steps:

[0007] Receive standardized order data, and combine it with a pre-established order causal network model to generate an initial state evaluation result and abnormal index data, and construct a state transition hypothesis set based on this;

[0008] Based on the order causal network model and the state transition hypothesis set, perform non-linear backtracking analysis to generate a state backtracking result;

[0009] Read the status backtracking result, execute the differential backtracking strategy, generate and execute an optimized backtracking plan, and update the order status;

[0010] Based on the status backtracking result and the updated order status, perform Bayesian network forward-backward inference fusion to generate the final status evaluation result.

[0011] Beneficial effects: The present invention breaks through the limitation of the preset status boundary, can flexibly handle various abnormal states and intermediate states, and improves the adaptability in complex business scenarios; it can accurately handle complex scenarios such as status bifurcation, merging, and cross-branch backtracking, and improves the accuracy and real-time performance of order status tracking in complex environments, especially performing well in the case of data delay and broken causal chains. Description of the Drawings

[0012] Figure 1 It is a step flowchart of a method for real-time tracking of recharge order status provided by an embodiment of the present application.

[0013] Figure 2 It is a step flowchart of constructing a state transition hypothesis set provided by an embodiment of the present application.

[0014] Figure 3 It is a step flowchart of performing non-linear backtracking analysis provided by an embodiment of the present application.

[0015] Figure 4 It is a step flowchart of performing a differential backtracking strategy provided by an embodiment of the present application. Detailed Embodiment

[0016] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0017] It should be specifically noted that, for clearly showing the step flow of the present application, numbers are marked for each step in the specification. These numbers are only for the convenience of description and do not limit the execution order of the steps. In actual operation, according to the technical requirements of specific implementation scenarios, the steps can be executed in an order different from that shown in the specification, and in some cases, parallel processing between steps can also be achieved.

[0018] As Figure 1 shown, a method for real-time tracking of recharge order status includes the following steps:

[0019] S1. Receive the standardized order data, and combine it with the pre-established order causal network model to generate the initial state evaluation result and abnormal index data, and construct a state transition hypothesis set based on this;

[0020] Specifically, the standardized order data is the standardized data of all recharge orders, such as order amount, time, user information, etc. The order causal network model can understand the causal relationship between order states. For example, if a certain order state becomes "failed", the system can analyze what factors caused it - whether it is a payment problem, a system failure, or a user operation error. Analyze the order data to determine which orders may have problems (such as failure or delay), obtain the initial state evaluation result and generate relevant abnormal indicators. In order to predict possible changes in the order, the system will establish different state change hypotheses. For example, it may assume that: "If a certain order is not processed for a long time, it may become an 'overtime' state."

[0021] S2. Based on the order causal network model and the state transition hypothesis set, perform non-linear backtracking analysis to generate a state backtracking result;

[0022] Specifically, for the situation where data arrives late or the state is inconsistent, the system will flexibly backtrack the order historical data instead of following a fixed order to find the key factors. For example, an order failure may not be caused by a single link, but by the combined effect of multiple factors. Non-linear backtracking analysis can help discover more complex relationships. Thus, a conclusion can be drawn, such as "The order failed because the payment system was slower during the peak period", providing a basis for subsequent optimization.

[0023] S3. Read the state backtracking result, execute a differential backtracking strategy, generate an optimized backtracking plan and execute it to update the order state;

[0024] Specifically, different orders may have different problems. Therefore, the system will not use the same method to solve all orders, but formulate different backtracking strategies according to the specific situation. For example: If an order fails due to a network problem, the system may automatically retry; if an order fails due to user payment failure, the system may send a reminder or suggest changing the payment method. After analyzing the order problem, the system will formulate the optimal solution and execute it immediately. For example, it may adjust the order processing sequence, optimize system resources, and reduce the failure rate.

[0025] S4. Based on the state backtracking result and the updated order state, perform Bayesian network forward-backward inference fusion to solve the state bifurcation and merging problem and generate the final state evaluation result.

[0026] Specifically, the forward-backward inference fusion of the Bayesian network uses probabilistic inference to predict and optimize decisions. It can infer the most likely future state based on the known order status (such as success, failure, processing). Forward inference can infer possible future changes from the current order status. For example, "If the order status is 'processing', is it most likely to become'success' or 'failure'?" Backward inference can trace back possible paths from the final goal. For example, "If we hope the order to be'successful' ultimately, what key links need to be ensured without errors?" Combining the analysis results of forward and backward inferences, we can obtain the most reasonable order status assessment. For example: This order is very likely to succeed and requires no additional processing; the order may fail, and it is recommended that the system take optimization measures (such as checking the payment method in advance); the order status is unstable, and the system needs to further monitor and adjust resource allocation.

[0027] This embodiment realizes efficient and intelligent order status tracking and optimization, which helps to improve system stability, reduce the exception rate, and enhance the user experience.

[0028] According to one aspect of the present application, the steps of generating an initial state assessment and constructing a set of state transition hypotheses include:

[0029] S11. Collect the original order data from different payment gateways, bank systems, and internal processing modules, perform standardization and noise filtering to obtain standardized order data;

[0030] S12. Based on the standardized order data and predefined business rules, construct an order causal network model representing the order status transition relationship as the basis for subsequent backtracking and prediction;

[0031] S13. Based on the standardized order data and the order causal network model, evaluate the current order status and detect potential anomalies, generating an initial state assessment result and anomaly metric data;

[0032] S14. When data delay or causal chain breakage is detected, based on the order causal network model and the anomaly metric data, generate multiple possible sets of state transition hypotheses to cope with uncertainties.

[0033] According to one aspect of the present application, the steps of generating an initial state assessment result and anomaly metric data include:

[0034] Read the standardized order data, extract the temporal characteristics, payment behavior characteristics, and system response characteristics of the order through the sliding time window analysis technique, and generate an order status feature vector;

[0035] Match the order status feature vector with the expected state transition path in the order causal network model, calculate the causal consistency score, and generate a causal consistency metric;

[0036] Based on the arrival situations of each data source in the standardized order data, combined with the key degrees of each data source in the order causal network model, calculate the completeness of the current status information, and generate an information completeness index;

[0037] Combined with the causal consistency index, the information completeness index, and the historical anomaly pattern library, apply an improved isolation forest algorithm to detect potential anomalies. This algorithm especially enhances the sensitivity to temporal anomalies and causal anomalies, and generates anomaly index data;

[0038] Based on the order status feature vector, the causal consistency index, and the anomaly index data, use a weighted voting mechanism to determine the current most likely status of the order, and at the same time calculate the confidence level of this determination, and generate an initial status evaluation result.

[0039] As Figure 2 shown, according to one aspect of the present application, the steps of constructing a state transition hypothesis set include:

[0040] Analyze the anomaly index data and the initial status evaluation result, identify the breakpoint location result of the causal chain in the order causal network model and perform historical pattern matching based on this, and generate a historical reference path set;

[0041] Combined with the breakpoint location result, construct a dynamic hypothesis space, and generate an extended state network;

[0042] Based on the extended state network and the historical reference path set, perform multi-scale probability hypothesis generation, obtain a preliminary state transition hypothesis set and optimize and prune it, and generate a final state transition hypothesis set.

[0043] Specifically, use a graph structure analysis algorithm to locate potential causal chain breakpoints in the order causal network model, and generate a breakpoint location result; perform similarity matching between the breakpoint location result and the historical breakpoint pattern library, extract the most similar historical cases and their solution paths, and generate a historical reference path set; based on the breakpoint location result, dynamically expand the state space in the order causal network model, create an extended network including additional intermediate states and abnormal branch states. This step breaks through the limitation of the fixed state space in traditional methods, and generates an extended state network; on the extended state network, combined with the historical reference path set and the currently observable state transition probability, use the Monte Carlo tree search method to generate multiple possible state transition paths, and assign probability weights to each path, and generate a preliminary state transition hypothesis set. Apply the information entropy maximization criterion to prune and optimize the state transition hypothesis set, retain the hypothesis paths with the highest information value, and at the same time ensure the diversity and representativeness of the hypotheses, and generate a final state transition hypothesis set.

[0044] According to one aspect of the present application, the steps of constructing a dynamic hypothesis space and generating an extended state network include:

[0045] Extract breakpoint information from the breakpoint positioning results, retrieve adjacent state nodes in the order causal network model, obtain a local network subgraph, and calculate the state transition logic difference value matrix between the breakpoints and adjacent nodes in it;

[0046] Generate candidate intermediate states at positions where the difference value exceeds a predetermined threshold according to the state transition logic difference value matrix;

[0047] Retrieve historical cases similar to the current breakpoint characteristics from the pre-stored historical abnormal pattern library, and extract non-standard state nodes;

[0048] Combine the candidate intermediate states and non-standard state nodes to construct a preliminary extended state set;

[0049] For the preliminary extended state set, merge the extended states with feature similarity exceeding the preset threshold and perform causal consistency verification to generate an extended state network.

[0050] According to one aspect of the present application, the steps of calculating the state transition logic difference value matrix between the breakpoints and adjacent nodes in the local network subgraph include:

[0051] Extract the multi-dimensional feature vectors of each node from the local network subgraph,

[0052] For each pair of adjacent nodes (i, j), calculate the similarity of the multi-dimensional feature vectors to generate a preliminary similarity value;

[0053] Adjust the preliminary similarity value according to the state transition time series data between nodes, increase the weight of recent transitions, and generate a time series weighted similarity;

[0054] Analyze the position and connection pattern of the nodes in the local network subgraph, adjust the calculation of the time series weighted similarity, and generate a structure fusion similarity;

[0055] Convert the structure fusion similarity into a difference value, and generate a state transition logic difference value matrix by using the method of 1 minus the similarity.

[0056] Specifically, extract breakpoint coordinate information from the breakpoint location results, including the state node ID where the breakpoint is located, the breakpoint timestamp, and the breakpoint type identifier. For each breakpoint, retrieve the set of adjacent state nodes in the order causal network model, including direct predecessor nodes and direct successor nodes, and construct a local network subgraph. Apply an improved graph structure expansion algorithm to expand the local network subgraph: calculate the state transition logic difference value matrix between the breakpoint and adjacent nodes; based on the difference value matrix, identify the positions most likely to require the insertion of intermediate states; for positions where the difference value exceeds a predetermined threshold (default set to 0.65), generate candidate intermediate states. Based on the historical anomaly pattern library, retrieve historical cases similar to the current breakpoint characteristics, and extract non-standard state nodes that appear in these cases. The similarity calculation uses the weighted cosine similarity method, giving higher weights to cases closer in time.

[0057] Combine the generated candidate intermediate states and non-standard state nodes in historical cases to construct a preliminary extended state set. For each extended state, assign its initialization parameters, including: state ID: in the form of the original breakpoint ID plus an extension identifier; state feature vector: generated from the feature vectors of adjacent standard states by interpolation; transition probability matrix: initialized as the weighted average of the transition probabilities of adjacent standard states. Apply a redundancy detection algorithm to the preliminary extended state set to merge extended states with a feature similarity exceeding 85%, avoiding unnecessary expansion of the state space. Integrate the refined extended state set with the original order causal network model, update the network topology structure, recalculate the connection relationships and weights between nodes, and generate a temporary extended network. Apply a network consistency verification algorithm to ensure that the extended network satisfies basic causal consistency rules, such as acyclicity, state reachability, etc. For parts that do not meet the rules, correct them by fine-tuning the connection weights or removing low-confidence connections. Finally, form a complete extended state network, which contains the original state nodes, newly extended intermediate state nodes, and updated transition relationships, providing a more complete state space for subsequent multi-path hypothesis generation.

[0058] This embodiment dynamically expands the state space, breaks through the limitation of the fixed state space in traditional methods, solves the problem that the traditional fixed state space model cannot handle unpreset abnormal states, enables the system to process undefined abnormal states in advance, improves the adaptability to complex scenarios, and enhances the system's ability to handle unknown abnormal patterns. In practical applications, when encountering unknown state transition patterns caused by situations such as bank system upgrades or new payment channel access, the system can intelligently identify and model these new states without manual intervention or system reconstruction. Actual measurements show that this embodiment increases the abnormal state recognition rate from 68% of traditional methods to 91%, reduces the demand for manual intervention by approximately 75%, and at the same time shortens the adaptation cycle of the system to new business scenarios from an average of 3 days to 4 hours, improving the business adaptability and operation and maintenance efficiency of the platform.

[0059] As Figure 3 shown, according to one aspect of the present application, the steps of performing non-linear backtracking analysis and generating a state backtracking result include:

[0060] S21. When obtaining new latency data, integrate it with the standardized order data, update the order state feature vector, generate an updated state feature, compare it with the initial state evaluation result, identify logical conflict points, and obtain a conflict recognition result;

[0061] S22. Based on the conflict recognition result and the order causal network model, construct multiple backtracking paths from the current state to possible historical states, that is, non-linear backtracking paths, and generate a backtracking path graph;

[0062] S23. Analyze the severity of the conflict recognition result and the current load condition of the system, determine the optimal backtracking depth, and generate a backtracking depth parameter;

[0063] S24. Verify the updated state feature with each assumed path in the state transition hypothesis set, calculate the matching degree score, and combine the backtracking path graph and the backtracking depth parameter to limit the verification range, obtain the hypothesis verification result and evaluate the impact, and accordingly select the optimal backtracking strategy to generate a state backtracking result.

[0064] Specifically, compare the updated status features with the previous initial status assessment results, identify potential logical conflict points through causal reasoning, which breaks through the limitation of simple time-series comparison in traditional linear backtracking, and generate conflict identification results. Multiple backtracking paths from the current state to possible historical states are not limited to simple linear backtracking, but allow complex backtracking across nodes and branches. Analyze the severity and scope of the conflict identification results, and dynamically determine the optimal backtracking depth in combination with the current load status of the system, which not only ensures the correction of necessary error states but also avoids wasting system resources due to excessive backtracking, and generate a backtracking depth parameter. Based on the hypothesis verification results, evaluate the impact of each possible backtracking operation on the current state of the order, the historical state chain, and related business processes, and generate a backtracking impact assessment report, which includes the impact scope, risk level, and processing complexity of each possible backtracking operation. Integrate the hypothesis verification results and the backtracking impact assessment report, select the optimal backtracking strategy, and generate a detailed status correction plan, which includes the historical state points to be modified, the modification content, and the execution order, and generate a status backtracking result.

[0065] According to one aspect of the present application, the steps of constructing a non-linear backtracking path and generating a backtracking path graph include:

[0066] Extract conflict point information from the conflict identification results, locate the corresponding state nodes in the order causal network model, construct a local network view, and search for potential state source points in the historical direction along the conflict points from it, calculate the probability score of each node as a conflict source, and combine the state nodes exceeding a predetermined threshold to generate a preliminary set of backtracking points;

[0067] Based on the preliminary set of backtracking points, construct possible backtracking paths from each backtracking point to the conflict point, allowing the paths to freely cross in the network considering implicit dependencies, and generate a preliminary set of paths;

[0068] Based on the preliminary set of paths, calculate the feasibility score for each path, combine the pre-stored path length, state transition probability, and historical similarity, generate a set of paths with scores, and perform filtering and priority sorting to obtain qualified paths to form a directed graph, and obtain a backtracking path graph after adding metadata.

[0069] According to one aspect of the present application, the steps of calculating the probability score of each node as a conflict source include:

[0070] Based on the local network view and conflict point information, initialize the search queue, set the conflict point as the search starting point, and analyze the connection relationships in the local network view, and assign conversion weights to each node, including forward and backward conversion weights, which respectively represent the strength of the cause and the result;

[0071] Determine the optimal search depth according to the severity and type of the conflict point information;

[0072] Extract similar conflict patterns from the historical conflict case library, calculate the frequency of each node as the source point in historical similar conflicts, and generate a historical weight vector.

[0073] Based on the transformation weight, the optimal search depth, and the historical weight vector, calculate the probability score of each node as the conflict source through an adaptive weighting formula.

[0074] Specifically, extract conflict point information from the conflict recognition results, including the conflict status node ID, conflict type identifier, timestamp, and conflict severity score (a floating point number in the range of 0 - 1). Locate the status node corresponding to each conflict point in the order causal network model, and construct a local network view centered on this node. This view includes the upstream and downstream nodes related to the conflict point and their connection relationships. Apply an improved bidirectional breadth-first search algorithm to search for potential status source points in the historical direction from the conflict point: the search is not limited to direct predecessor nodes, but is prioritized based on the strength of causal relationships; each time the search depth is extended, both the direct connections and indirect influence paths between nodes are considered; calculate the probability score of each extended node as the conflict source based on node characteristics and historical patterns. Generate a preliminary set of backtracking points, including all status nodes whose probability scores exceed the threshold (default is 0.4).

[0075] Apply a path generation algorithm to construct possible backtracking paths from the conflict point to each candidate backtracking point: different from traditional linear backtracking, this algorithm allows paths to freely traverse the network, considering implicit dependencies and cross-branch relationships; for cross-branch jumps, introduce a crossing cost coefficient to make the path tend to the optimal solution while maintaining flexibility; calculate the feasibility score for each generated path, combining path length, state transition probability, and historical similarity.

[0076] Apply a multi-dimensional filtering algorithm to the generated paths to eliminate low-feasibility paths. This algorithm comprehensively considers the following factors: the causal consistency score of the path: evaluate whether the path conforms to the causal relationships of business logic; the data support degree of the path: evaluate whether there is sufficient data to support the existence of this path; the historical frequency of the path: evaluate the frequency of this path in historical similar scenarios.

[0077] Sort the remaining paths by priority to generate a set of weighted paths. The weight calculation considers the path feasibility score, conflict severity, and business impact scope. Integrate all qualified paths into a directed graph structure to form a complete backtracking path graph. This graph is not a simple linear chain or tree, but a graph that allows complex topological structures, possibly including advanced structures such as multiple entrances, multiple exits, loops, and cross-edges. Add metadata to each edge and node in the backtracking path graph, including information such as transition probability, data dependency, and operation cost, to facilitate the formulation of subsequent backtracking strategies.

[0078] This embodiment breaks through the limitations of traditional linear backtracking, realizes backtracking based on the causal network's complex topological structure, can handle complex scenarios such as state forking, merging, and cross-branch backtracking, and improves the accuracy and flexibility of backtracking in complex order processing flows. Compared with traditional linear backtracking methods, in this embodiment, when dealing with multi-channel parallel payment and partial liquidation scenarios, the state backtracking accuracy rate has increased by 23 percentage points (from 72% to 95%). Especially in scenarios where data is severely delayed (>120 seconds) or partially permanently lost, the state judgment error rate has decreased by 68%, the average backtracking path length has decreased by 37%, and the processing complexity has dropped from O(n) to O(log n), ensuring that state backtracking analysis can be completed within 200 milliseconds even in the most complex order processing network, meeting the strict time requirements of real-time business processing.

[0079] Existing methods usually adopt a "one-size-fits-all" backtracking strategy, which fails to perform differential processing based on business importance and backtracking cost, may not only cause waste of system resources but also be difficult to ensure the continuity of critical business processes. Therefore, as Figure 4 shown, according to one aspect of the present application, the steps of executing a differential backtracking strategy and generating an optimized backtracking scheme include:

[0080] S31. Read the state backtracking result, calculate the backtracking cost of each backtracking operation, and generate a backtracking cost matrix;

[0081] S32. Analyze the business scenarios and user experience impacts involved in the state backtracking result, combine with predefined business priority rules, classify the priorities of backtracking operations, and generate business priority tags; formulate differential processing strategies for different types of backtracking operations in combination with the backtracking cost matrix, and generate a differential backtracking strategy table;

[0082] S33. According to the differential backtracking strategy table and the current system load status, optimize the execution order and resource allocation of backtracking operations, ensure that critical backtracking operations obtain sufficient resources, and at the same time avoid causing excessive pressure on the normal operation of the system, and generate a resource scheduling plan;

[0083] S34. Integrate the differential backtracking strategy table, the resource scheduling plan, and the business priority tags to generate an optimized backtracking scheme.

[0084] Specifically, read the status backtracking results and apply a multi-factor cost evaluation model, which considers dimensions such as computing resource consumption, business impact scope, status recovery difficulty, and time sensitivity, to calculate a comprehensive cost value for each backtracking operation and generate a backtracking cost matrix. Based on the backtracking cost matrix and business priority tags, formulate differentiated processing strategies for different types of backtracking operations: for high-priority and low-cost operations, implement immediate full-scale backtracking; for high-priority and high-cost operations, implement phased progressive backtracking; for low-priority operations, implement delayed backtracking or simplified backtracking to generate a differentiated backtracking strategy table. Convert the differentiated backtracking strategy table and resource scheduling plan into a series of atomic-level execution instructions, where each instruction includes a clear operation object, operation type, preconditions, and rollback mechanism to ensure the reliable execution of backtracking operations and the rollback ability when necessary, and generate backtracking execution instructions. Integrate the differentiated backtracking strategy table, resource scheduling plan, and execution priority to generate a complete optimized backtracking plan, which includes a comprehensive execution plan, monitoring point settings, and emergency handling mechanisms.

[0085] According to one aspect of the present application, the steps of generating a differentiated backtracking strategy table include:

[0086] Map the backtracking operations in the backtracking cost matrix to the 0 to 1 interval according to the cost value and divide them into three cost levels;

[0087] Classify the operations in the business priority tags into three priority levels according to business importance;

[0088] Construct a strategy mapping matrix with the cost level as the row and the priority level as the column; where in the strategy mapping matrix, each cell contains a processing strategy template applicable to the combination;

[0089] According to the position of each backtracking operation in the strategy mapping matrix, assign a preliminary strategy type, and for each preliminary strategy type, refine the specific execution parameters to obtain a preliminary strategy; where the execution parameters include execution priority, execution delay, and backtracking data range; the preliminary strategy types include immediate full-scale backtracking, phased progressive backtracking, and delayed backtracking or simplified backtracking;

[0090] According to the system load condition, time urgency, and operation dependency relationship, make fine-grained adjustments to the preliminary strategy and merge similar backtracking operations to obtain an adjusted strategy;

[0091] Based on the adjusted strategy, generate an execution specification for each backtracking operation and integrate the execution specifications of all operations to obtain a differentiated backtracking strategy table. Where the execution specification includes execution timing, execution method, resource limitations, and monitoring points.

[0092] According to one aspect of the present application, the steps of constructing a strategy mapping matrix include:

[0093] Extract the basic backtracking policy template from the system preset policy library; it includes three basic policies: immediate full volume, phased progressive, and delayed simplification.

[0094] Perform a fine-grained division of the cost level to obtain a refined cost level; among the three cost levels of high, medium, and low, each level is further divided into three sub-levels, forming a nine-level refined cost level.

[0095] According to the current business scenario data, apply context-sensitive enhancement to the priority level, considering time factors, user influence scope, and business relevance, to generate an enhanced priority level.

[0096] Based on the refined cost level, enhanced priority level, and basic backtracking policy template, construct an initial policy mapping matrix.

[0097] Detect the mutation points between adjacent policies in the initial policy mapping matrix, identify policy conflicts that may lead to uneven resource allocation or sudden changes in execution efficiency, generate a conflict resolution matrix and perform smoothing processing to generate a policy mapping matrix.

[0098] Specifically, classify each backtracking operation in the backtracking cost matrix according to the cost value, map the cost value to the interval [0, 1], and divide it into three cost levels of high, medium, and low, with the demarcation points set at 0.7 and 0.4 respectively. Similarly, classify each operation in the business priority label into three levels: critical, important, and general according to business importance. Construct a 9x9 policy mapping matrix, where the rows represent the cost levels (high, medium, low), the columns represent the business priorities (critical, important, general), and each cell contains the specific processing policy template applicable to this combination.

[0099] For each backtracking operation, allocate a preliminary policy type according to its position in the policy mapping matrix: for operations with critical business and low cost, allocate the "immediate full volume backtracking" policy; for operations with critical business but high cost, allocate the "phased progressive backtracking" policy; for operations with general business and high cost, allocate the "delayed backtracking or simplified backtracking" policy; other combinations adopt corresponding intermediate policies.

[0100] For each policy type, further refine the specific execution parameters: for the "immediate full volume backtracking" policy, set the highest execution priority, no execution delay, and backtrack all data; for the "phased progressive backtracking" policy, design a multi-stage execution plan: the first stage: immediately backtrack the critical data fields with medium-high priority; the second stage: backtrack the important data fields when the system load is low with medium priority; the third stage: backtrack the remaining data fields during the maintenance window period with low priority. For the "delayed backtracking or simplified backtracking" policy, set a low execution priority, allow it to be executed when the system is idle, or only backtrack the necessary status flag fields.

[0101] Apply the context - sensitive adjustment algorithm to perform fine - grained adjustment on the preliminary strategy according to the current system load status, time urgency, and the dependency relationships of associated operations: Consider the dependency relationships between operations to ensure that prerequisite operations are executed first; Consider time sensitivity and increase the priority of operations approaching timeout; Consider resource conflicts and stagger high - cost operations that require the same resources.

[0102] Apply the batch processing optimization algorithm to identify similar rollback operations that can be batch - processed and merge them into batch tasks to improve efficiency. This algorithm uses an improved clustering method to group according to operation type, target data area, and resource requirements. For each rollback operation or batch task, generate detailed execution specifications, including: Execution timing: immediate, delayed, scheduled, or conditionally triggered; Execution method: full - volume, incremental, or differential; Resource limits: maximum CPU usage, memory occupancy, and IO bandwidth; Rollback conditions: define the rollback trigger conditions and rollback process when execution fails; Monitoring points: set monitoring points and alarm thresholds for key execution phases. Integrate the execution specifications of all operations to generate a complete differential rollback strategy table, which contains the detailed execution plans, resource requirements, and priority arrangements of each rollback operation.

[0103] This embodiment realizes differential rollback processing based on cost and business impact, avoids the resource waste and business interruption risks brought by the "one - size - fits - all" rollback in traditional methods, and improves the system resource utilization efficiency and business continuity. In the production environment test, this embodiment increases the system processing capacity during peak periods by 35%, reduces the CPU usage by 20%, and reduces the memory occupancy by 15%. Especially when dealing with a large number of order exceptions, by delaying low - priority operations to periods with lower system load, the response time of critical services is shortened by 67%, and at the same time, 85% of business interruption events are reduced. Further analysis shows that this embodiment improves the system resource utilization efficiency by 42%, while ensuring the real - time nature of critical services, reduces system fluctuations caused by resource competition, and provides users with a more stable and predictable service experience.

[0104] According to one aspect of the present application, the steps of performing Bayesian network forward - backward inference fusion to generate the final state evaluation result include:

[0105] S41. Extract historical state evidence and real - time state evidence from the state rollback result and the updated order state to generate a two - way evidence set;

[0106] S42. Convert the two - way evidence set into the probability distribution of nodes in the two - way Bayesian network to generate an evidence assignment result with confidence; at the same time, transmit from the past state to the future state and from the current observation to the historical state to generate a joint probability distribution;

[0107] S43. Adjust the joint probability distribution to satisfy the consistency rule constraints, generate the constrained adjusted probability distribution, extract state information at different granularities from it, generate multi-granularity state inference results, and combine the optimized backtracking scheme and the latest observed evidence to generate the final state evaluation result.

[0108] Specifically, read the state backtracking result and update the state features, extract evidence variables from two directions: the historical state evidence provided by backtracking and the real-time state evidence provided by the current observation, and generate a two-way evidence set. Dynamically construct a Bayesian network structure for the current order state based on the order causal network model. This network is specially designed with a node connection topology that supports two-way reasoning, breaking through the limitation of the one-way reasoning of traditional Bayesian networks, and generating a two-way Bayesian network. Convert the evidence in the two-way evidence set into the probability distribution or deterministic evidence of the nodes in the two-way Bayesian network according to reliability and timeliness. For uncertain evidence, retain its uncertainty instead of simply assigning values, which breaks the limitation that evidence must be certain in traditional Bayesian reasoning, and generate an evidence assignment result with credibility. Implement an improved message passing algorithm on the two-way Bayesian network, allowing evidence influence to be transmitted simultaneously from the past state to the future state (forward) and from the current observation to the historical state (backward). This algorithm specifically strengthens the ability to handle cyclic dependencies and state fork merges, and generates a joint probability distribution. Convert the consistency rules in the business logic into hard constraints or soft constraints in the two-way Bayesian network, adjust the joint probability distribution to satisfy these constraints, solve the complex business rule problems that are difficult to handle by traditional Bayesian networks, and generate the constrained adjusted probability distribution. Extract state information at different granularities from the constrained adjusted probability distribution, including deterministic high-level state judgments and fine-grained state distributions that retain uncertainty. This multi-granularity output breaks through the limitation of black-and-white state judgments in traditional methods, and generates multi-granularity state inference results. Integrate the multi-granularity state inference results, the optimized backtracking scheme and the latest observed evidence to generate the final state evaluation result, which not only includes the current most likely order state, but also includes the confidence interval of the state judgment, the possible change trend and the uncertainty measure.

[0109] In the case of processing parallel operations of multiple payment channels, traditional forward or backward one-way reasoning methods cannot effectively integrate the historical state path and the current observed evidence, and it is difficult to provide a reliable state evaluation result, seriously affecting the platform's ability to timely intervene and process abnormal orders. Therefore, according to one aspect of the present application, the steps of generating the joint probability distribution include:

[0110] Perform two-way topological sorting on the two-way Bayesian network to generate a node processing sequence; combine the evidence assignment result with credibility to generate an initial message set;

[0111] Based on the initialization message set, perform forward pass in the order of topological sorting to generate a forward message set; perform backward pass in the order of reverse topological sorting to generate a backward message set; process the forward and backward messages for each node simultaneously to update the node belief state; and process the cyclic dependence structure in the network to generate a revised belief state.

[0112] Based on the revised belief state, apply special processing rules to state bifurcation and merger points to generate a further revised belief state.

[0113] Monitor the change rate of the further revised belief state in the network. Stop the iteration when the global change rate is lower than a predetermined threshold or the maximum number of iterations is reached, and extract the joint probability distribution from the final belief state of each node.

[0114] Specifically, convert the result of evidence assignment with belief into the initial belief state of the corresponding nodes in the network, including both deterministic evidence and evidence in the form of probability distribution. Perform topological sorting on the bidirectional Bayesian network to generate a node processing sequence. Different from traditional algorithms that only consider the one-way sorting from cause to effect, here an improved bidirectional topological sorting algorithm is adopted, which can support both forward and backward information flows and can consider both causal directions and reverse dependencies simultaneously.

[0115] Construct and implement an innovative bidirectional message passing protocol: Define two types of messages for each network connection: forward message (from cause to effect) and backward message (from effect to cause); the message content includes probability distribution and uncertainty assessment, represented by Dirichlet distribution parameters; each message is attached with an information flow source identifier to prevent feedback loops. Initialize the messages of all network connections to a uniform distribution or a specified distribution, indicating a state of no information.

[0116] Apply the improved Sum-Product algorithm for message passing: In the forward pass stage: In the order of topological sorting, each node collects all input messages, integrates its own evidence, and then sends the updated forward message to all successor nodes; in the backward pass stage: In the order of reverse topological sorting, each node collects all backward messages, integrates the backward evidence, and then sends the updated backward message to all predecessor nodes; in the bidirectional integration stage: Each node processes the forward and backward messages simultaneously, resolves possible conflicts, and updates the node belief state.

[0117] Process the cyclic dependence structure in the network, which is a problem difficult to handle in traditional Bayesian networks: Implement an improved Loopy Belief Propagation algorithm, specifically for dealing with loop structures; introduce a message decay factor to prevent information from being over-amplified during cyclic propagation; set the maximum number of iterations and convergence conditions to ensure the stability of the algorithm in complex networks.

[0118] Special logic for handling state bifurcation and merging points: At the bifurcation point, where one node affects multiple successor nodes, specialized information distribution rules are implemented to ensure that the information received by different branches conforms to their causal association strength; at the merging point, where multiple predecessor nodes affect one node, weighted information fusion rules are implemented to integrate evidence based on the reliability and relevance of each predecessor node.

[0119] Apply a dynamic convergence judgment mechanism to monitor the change rate of the belief state in the network. Stop iteration when the global change rate is lower than a threshold (default is 0.01) or reaches the maximum number of iterations (default is 20). After convergence, extract the comprehensive probability distribution from the final belief state of each node to form the joint probability distribution of the entire network, which reflects the globally consistent state estimate after fusing forward and backward inferences.

[0120] This embodiment unifies forward and backward inferences into the same Bayesian network framework, realizing the fusion of the ability to correct historical states based on existing observations and predict future states based on historical paths, solving the problem of the separation of forward and backward inferences in traditional methods, and particularly enhancing the ability to handle circular dependencies and state bifurcation and merging. Practical application data shows that compared with traditional one-way inference methods, this embodiment increases the accuracy of state inference by 18 percentage points (from 79% to 97%). Especially in dealing with complex order scenarios with loop paths and multi-branch merges, the misjudgment rate drops from 22% to 3%. Importantly, the system can give accurate pre-judgments in 85% of cases before all relevant data arrives, with an average pre-judgment leading time of 43 seconds, providing a valuable lead for business intervention and exception handling, and improving the system's response speed and handling efficiency for abnormal situations.

[0121] According to one aspect of the present application, the steps of extracting state information at different granularities and generating multi-granularity state inference results include:

[0122] Extract the marginal probability distribution of each state node from the probability distribution after constraint adjustment, and aggregate the probability distributions of fine-grained state nodes into a multi-level state aggregation result;

[0123] Calculate the uncertainty measure for the state aggregation result of each level to generate a state uncertainty index; combine the risk level of the current business scenario to determine the optimal decision threshold for each level;

[0124] Generate multi-level deterministic state results based on the optimal decision threshold and probability distribution; for low-determinacy cases among them, generate auxiliary decision-making information;

[0125] Retain the complete probability distribution and uncertainty measure for internal use by the system to generate internal complete state information;

[0126] Generate appropriate state representations for different interfaces according to the recipient's requirements and processing capabilities;

[0127] Integrate multi-level state aggregation results, auxiliary decision-making information, internal complete state information, and state representations to generate a complete multi-granularity state inference result.

[0128] Specifically, extract the marginal probability distribution of each state node from the constrained adjusted probability distribution to generate a state probability vector at the node level. Apply the hierarchical state aggregation algorithm. According to the predefined state hierarchy system, aggregate the probability distributions of fine-grained state nodes into higher-level state categories: Basic layer: Retain the complete probability distribution of the original fine-grained state nodes; Business layer: Map the technical state to the main state categories defined by the business; Decision layer: Further aggregate into the core state categories to support decision-making.

[0129] For the state aggregation results of each level, calculate the uncertainty measures: Use Shannon Entropy to calculate the uncertainty of the probability distribution; Use the Diversity Index to evaluate the dispersion degree of the distribution; Use the ratio of the highest probability to the second-highest probability to calculate the certainty strength of the state judgment.

[0130] Apply the improved decision threshold dynamic adjustment algorithm to determine the optimal decision threshold for each level according to the current business scenario and uncertainty measures: High-risk business scenario: Require a higher certainty threshold, default set to 0.85; General business scenario: Adopt a medium certainty threshold, default set to 0.7; Low-risk scenario: Adopt a lower certainty threshold, default set to 0.6.

[0131] Based on the decision threshold and probability distribution, generate multi-level certainty state judgments: High certainty judgment: When the highest probability state exceeds the threshold, directly output this state as the certainty result; Medium certainty judgment: When the highest probability is close to but does not exceed the threshold, output this state but attach an uncertainty warning; Low certainty judgment: When the probability distribution is relatively dispersed, output multiple possible states and their probabilities without giving a definite judgment.

[0132] For low-certainty situations, apply a decision support enhancement algorithm to provide auxiliary information: generate key observation suggestions indicating which additional information may improve judgment certainty; calculate the conditional expected risk to evaluate the risks of making different decisions under current uncertainties; provide risk mitigation measure suggestions for high-risk decision points. For the state information used internally in the system, retain the complete probability distribution and uncertainty metrics to support probabilistic reasoning and risk assessment in subsequent processing steps. For the state information presented to external systems or user interfaces, provide state representations with different granularities according to the requirements and processing capabilities of the recipients: Technical interface: Provide the complete probability distribution and uncertainty metrics; Business interface: Provide the main state categories and their probabilities; User interface: Provide a simplified certainty state and represent uncertainty through visual elements (such as color coding). Integrate the state information at all these levels and representation forms to generate a complete multi-granularity state inference result, which contains multi-dimensional state descriptions from details to overviews and from certainty to uncertainty.

[0133] This embodiment provides state judgment results with different certainty levels, which not only meet the business scenarios that require clear decisions but also provide uncertainty quantification for risk assessment and early warning, enhancing the practicality and interpretability of the system. This embodiment improves the decision-making accuracy by 27% and simultaneously increases the accuracy rate of risk early warning by 39%. Data analysis shows that in high-uncertainty scenarios (such as new payment channels or abnormal network environments), by retaining and quantifying uncertainties, the system enables the risk control module to make correct intervention decisions in over 90% of cases, reducing false alarms and missed alarms by 76%. Meanwhile, providing customized state expressions for different business interfaces improves the processing efficiency of downstream systems by 25% and reduces the data parsing error rate by 82%, enhancing the collaborative efficiency and interoperability of the entire payment ecosystem.

[0134] According to one aspect of the present application, it further includes: S5. Based on the actual order status conversion result, dynamically update and optimize the order causal network model to improve the system's adaptability to new models, and generate an updated network parameter and learning result report.

[0135] In a specific embodiment of the present application, a real-time tracking method for recharge order status based on a causal network-driven non-linear backtracking algorithm is described. It can be applied to large electronic payment platforms to process recharge orders of multiple payment channels (bank cards, e-wallets, third-party payments, etc.). The system needs to track the status of each order in real time and make accurate status judgments and backtracking updates in case of data delay, loss, or inconsistency. In this embodiment, the system receives order data for a recharge via a bank card. The order number is ORDER2025032900123, the user ID is USER98765, the recharge amount is 1000 yuan, and the recharge time is 2025-03-29 14:30:25. The specific steps are as follows:

[0136] Step 1: Conduct real-time status evaluation and anomaly detection.

[0137] 1.1 Extract multi-dimensional status features.

[0138] The system reads the standardized order data, including the following key fields: Order basic information: order number, user ID, amount, creation time; Payment gateway data: gateway response code, authorization timestamp, transaction reference number; Bank system data: processing status code, clearing flag, callback timestamp; Internal processing data: account verification status, balance update flag, notification sending status.

[0139] The system uses a 5-second sliding time window to extract the following features: Temporal features: state transition intervals, cumulative processing time, response delay patterns; Payment behavior features: matching degree of user's historical payment patterns, amount features, channel characteristics; System response features: response code patterns of each processing node, processing duration distribution. These features are organized into a 92-dimensional order status feature vector F, where F[1:30] is the temporal feature sub-vector; F[31:60] is the payment behavior feature sub-vector; F[61:92] is the system response feature sub-vector.

[0140] 1.2 Conduct status evaluation and anomaly detection.

[0141] The system completes a preliminary evaluation based on the order status feature vector and the order causal network model to generate an initial status evaluation result. For the example order, the system initially determines its current status as "Payment authorization successful, waiting for clearing", the status code is S_AUTH_SUCCESS, and the confidence level is 0.89. At the same time, the system detects an anomaly: the bank clearing callback data is expected to arrive within 45 seconds, but it has been waiting for 75 seconds and still has not been received. Anomaly metric data is generated, the anomaly type is marked as "Clearing callback delay", the anomaly code is E_CLEARING_DELAY, and the severity is 0.6 (medium).

[0142] Step 2: Generate multi-path hypotheses.

[0143] 2.1. Identify the breakpoints of the causal chain.

[0144] The system analyzes the abnormal indicator data and the initial state assessment results, and locates the breakpoints in the order causal network model. For the example order, the breakpoint is located on the transition path from "Payment authorization successful" to "Clearing confirmation", the node ID is N_CLEARING_CONFIRM, the breakpoint type is "Data missing", and the breakpoint location result is generated.

[0145] 2.2. Construct a dynamic hypothesis space.

[0146] The system reads the breakpoint location result and the order causal network model, and executes the dynamic hypothesis space construction process: Extract adjacent nodes from the breakpoint N_CLEARING_CONFIRM, including the predecessor node N_AUTH_SUCCESS and the successor nodes N_BALANCE_UPDATE, N_FAIL_REFUND. Calculate the state transition logic difference value matrix D between the breakpoint and the adjacent nodes. For each pair of adjacent nodes (i, j), the difference value is calculated as: D(i, j) = 1 - cos(θ(Vi, Vj)); where Vi and Vj are the feature vectors of nodes i and j respectively, and cos(θ) represents the cosine similarity. The specific calculation is: D(N_AUTH_SUCCESS, N_CLEARING_CONFIRM = 0.72 D(N_CLEARING_CONFIRM, N_BALANCE_UPDATE) = 0.45 D(N_CLEARING_CONFIRM, N_FAIL_REFUND) = 0.81. According to the difference value matrix, the system determines that two positions need to insert intermediate states: between the predecessor node N_AUTH_SUCCESS and the breakpoint N_CLEARING_CONFIRM (difference value 0.72 > threshold 0.65); between the breakpoint N_CLEARING_CONFIRM and the successor node N_FAIL_REFUND (difference value 0.81 > threshold 0.65).

[0147] The system retrieves the historical anomaly pattern library and discovers 3 similar cases, with similarity degrees to the current breakpoint being 0.85, 0.72, and 0.63 respectively. It extracts the non-standard status nodes that appear in these cases: N_CLEARING_PENDING: the status of in-clearing process; N_CLEARING_RETRY: the status of clearing retry; N_PARTIAL_CONFIRM: the status of partial confirmation. The system constructs an extended status set, including: N_CLEARING_CONFIRM_EXT1: the feature vector is generated by linear interpolation of N_AUTH_SUCCESS and N_CLEARING_CONFIRM; N_CLEARING_CONFIRM_EXT2: the feature vector is generated by linear interpolation of N_CLEARING_CONFIRM and N_FAIL_REFUND; N_CLEARING_PENDING: the status extracted from historical cases; N_CLEARING_RETRY: the status extracted from historical cases; N_PARTIAL_CONFIRM: the status extracted from historical cases.

[0148] The system applies the redundancy detection algorithm and discovers that the feature similarity between status N_CLEARING_CONFIRM_EXT1 and status N_CLEARING_PENDING is 0.92 (>0.85 threshold), so they are merged into status N_CLEARING_PENDING. The system updates the network topology, adds new nodes and connections, and adjusts the transition probability to ensure network consistency. The finally generated extended status network contains newly added intermediate status nodes, enhancing the expressive power of the original network, especially the coverage of abnormal paths.

[0149] 2.3. Generate multi-path hypotheses.

[0150] Based on the extended state network, the system generates multiple state transition hypothesis paths, including: Hypothesis Path 1: N_AUTH_SUCCESS → N_CLEARING_PENDING → N_CLEARING_CONFIRM → N_BALANCE_UPDATE (probability 0.65); Hypothesis Path 2: N_AUTH_SUCCESS → N_CLEARING_PENDING → N_CLEARING_RETRY → N_CLEARING_CONFIRM (probability 0.25); Hypothesis Path 3: N_AUTH_SUCCESS → N_CLEARING_PENDING → N_PARTIAL_CONFIRM → N_BALANCE_UPDATE (probability 0.08); Hypothesis Path 4: N_AUTH_SUCCESS → N_CLEARING_CONFIRM_EXT2 → N_FAIL_REFUND (probability 0.02). Apply the criterion of maximizing information entropy to optimize the hypothesis set, and retain the final state transition hypothesis set, including Hypothesis Paths 1, 2, and 3.

[0151] Step 3. Perform non-linear backtracking analysis driven by the causal network.

[0152] 3.1. Perform delayed data integration and conflict detection.

[0153] During the waiting period of the system, delayed bank clearing callback data is received, indicating that the order has been partially confirmed (the amount is 700 yuan, and the remaining 300 yuan needs additional verification). The system integrates this new data and updates the order status feature vector to the updated status feature. The system compares the updated status feature with the initial state assessment result and discovers a logical conflict: the initial state assessment is waiting for full clearing, but actually partial clearing data is received. The system generates a conflict identification result, marking the conflict point as "clearing amount mismatch", and the conflict severity as 0.7.

[0154] 3.2. Construct non-linear backtracking paths.

[0155] The system constructs non-linear backtracking paths based on the conflict identification result and the order causal network model: The system starts from the conflict point and constructs a local network view in the order causal network model, including all state nodes related to clearing.

[0156] Apply the improved bidirectional breadth-first search algorithm to find potential state source points: For each candidate node n, calculate its probability P(n) as the conflict source: P(n) = α·Hist(n) + β·Caus(n) + γ·Time(n); where Hist(n) is the normalized frequency of node n appearing in historical similar conflicts; Caus(n) is the causal association strength between node n and the conflict point; Time(n) is the correlation weight based on the time order; α, β, γ are weight coefficients, satisfying α + β + γ = 1. For the example order, the weights are set as α = 0.3, β = 0.5, γ = 0.2, and the calculation results are: P(N_AUTH_SUCCESS) = 0.58; P(N_PAYMENT_INIT) = 0.35; P(N_ACCOUNT_VERIFY) = 0.42. Where P(N_AUTH_SUCCESS) is the probability of the order success authentication state; P(N_PAYMENT_INIT) is the probability of the order payment initialization state; P(N_ACCOUNT_VERIFY) is the probability of the account verification state as the conflict source. The system generates a preliminary set of backtracking points, including the predecessor node N_AUTH_SUCCESS and the account verification node N_ACCOUNT_VERIFY (probability value > 0.4 threshold).

[0157] The system constructs possible backtracking paths for each backtracking point and calculates the path feasibility score S_path using the following formula: S_path = ∏P(ti)·[1 -λ·Jumps]; where ∏P(ti) is the product of all transition probabilities on the path; Jumps is the number of cross-branch jumps in the path; and λ is the crossing cost coefficient, with a default value of 0.15. For the example order, the system generates the following backtracking paths: Path 1: N_AUTH_SUCCESS → N_CLEARING_PENDING → N_PARTIAL_CONFIRM, with a feasibility score of 0.65·0.75·[1 -0.15·0] = 0.4875; Path 2: N_ACCOUNT_VERIFY→ N_PAYMENT_INIT → N_AUTH_SUCCESS → N_CLEARING_PENDING, with a feasibility score of 0.85·0.92·0.65·[1 -0.15·1] = 0.4179; Path 3: N_AUTH_SUCCESS → N_CLEARING_PENDING→ N_CLEARING_RETRY → N_PARTIAL_CONFIRM, with a feasibility score of 0.65·0.35·0.60·[1 -0.15·0] = 0.1365. The system applies a multi-dimensional filtering algorithm to retain Path 1 and Path 2 and generates a weighted backtracking path graph.

[0158] 3.3. Perform backtracking depth control and result generation.

[0159] Based on the conflict severity (0.7) and system load (currently 45%), the system calculates the optimal backtracking depth to be 2 and generates a backtracking depth parameter. The updated state features are verified in parallel with the set of state transition hypotheses, and the best hypothesis path is determined to be Hypothesis Path 1 (matching degree 0.88) according to the matching degree. Evaluate the backtracking impact, generate a backtracking impact assessment report, and finally determine the state backtracking result: The order status needs to be backtracked and updated from "Payment authorization successful, waiting for clearing" to "Partial clearing confirmation", and the corresponding business logic is triggered.

[0160] Step 4. Execute the differential backtracking strategy.

[0161] 4.1. Generate differential strategies.

[0162] The system reads the state backtracking results and applies a multi-factor cost evaluation model to calculate the comprehensive cost C of the backtracking operation: C = w1·Ccomp + w2·Cbus + w3·Crecov + w4·Ctime; where Ccomp is the computing resource consumption (range 0-1); Cbus is the business impact scope (range 0-1); Crecov is the difficulty of state recovery (range 0-1); Ctime is the time sensitivity (range 0-1); w1, w2, w3, w4 are weight coefficients, satisfying ∑wi = 1. For the backtracking operation of the sample order, the parameter values are: Ccomp = 0.4 (medium computing consumption); Cbus = 0.7 (large business impact); Crecov = 0.3 (low recovery difficulty); Ctime = 0.8 (high time sensitivity); weights w1 = 0.2, w2 = 0.4, w3 = 0.1, w4 = 0.3. The calculated comprehensive cost C = 0.2×0.4 + 0.4×0.7 + 0.1×0.3 + 0.3×0.8 = 0.63, which belongs to the medium cost. According to the business scenario analysis, it is determined that this order has an "important" business priority, and a business priority mark is generated.

[0163] Map the cost (0.63, medium) and business priority (important) to the 9×9 strategy matrix, determine the preliminary strategy type as "phased progressive backtracking", and refine the execution parameters: First stage: Immediately backtrack the status flag and amount field, with a priority of 8 (high); Second stage: Backtrack the transaction history within 15 minutes, with a priority of 5 (medium); Third stage: Backtrack the associated data when the system load is below 30%, with a priority of 3 (low). Considering the current context, apply an adjustment algorithm to optimize the strategy, and finally generate a differentiated backtracking strategy table.

[0164] 4.2. Perform resource scheduling and execution.

[0165] According to the differentiated backtracking strategy table and the current system load (45%), optimize the resource allocation, generate a resource scheduling plan and backtracking execution instructions, and finally form an optimized backtracking plan.

[0166] Step Five. Perform forward-backward inference fusion of the Bayesian network.

[0167] 5.1. Perform evidence collection and allocation.

[0168] The system extracts two-way evidence from the state backtracking results and updated state features: Backtracking evidence: indicating that the order has undergone partial liquidation confirmation; Current observation: indicating that the remaining amount requires additional verification. Construct a two-way Bayesian network, allocate the evidence, and generate an evidence allocation result with credibility.

[0169] 5.2. Perform forward-backward message passing.

[0170] The system implements an improved Sum-Product algorithm on a two-way Bayesian network: First, perform a two-way topological sorting on the network to generate a node processing sequence. Initialize all connected messages to a uniform distribution. Conduct message passing iterations, and each iteration includes three stages: forward passing, backward passing, and two-way integration: For forward passing, each node X calculates the message M_X→Y sent to its child node Y: M_X→Y(y) = ∑_x P(Y=y|X=x)·[P(X=x)·∏_{Z∈Pa(X)\Y} M_Z→X(x)]; where P(Y=y|X=x) is the transition probability in the conditional probability table; P(X=x) is the local evidence of node X (if any); Pa(X) is the set of parent nodes of X; M_Z→X(x) is the message received by X from other parent nodes Z; ∏ is the product function; x is the possible value of the parent node X, and y is the possible value of the child node Y. For backward passing, each node Y calculates the message M_Y→X sent to its parent node X: M_Y→X(x) = ∑_y P(Y=y|X=x)·[P(Y=y)·∏_{Z∈Ch(Y)\X} M_Z→Y(y)]; where Ch(Y) is the set of child nodes of Y. For two-way integration, each node X updates its belief state Bel(X): Bel(X=x) =α· P(X=x)·∏_{Z∈Pa(X)} M_Z→X(x)·∏_{Y∈Ch(X)} M_Y→X(x); where α is the normalization constant.

[0171] Apply the improved Loopy Belief Propagation algorithm to handle cyclic dependency structures: Introduce a message decay factor d = 0.9; for the message M t _X→Y from node X to Y, if there is a cycle, then: M t _X→Y = d·M t _X→Y +(1-d)·M (t-1) _X→Y. The system applies special processing rules to state fork and merge points: Fork point: Node X affects multiple child nodes Y1, Y2,..., Yn, adjust the message passing weights; Merge point: Multiple parent nodes X1, X2,..., Xm affect node Y, use weighted information fusion. Monitor the change rate of the belief state, and stop when the global change rate is lower than 0.01 or reaches 20 iterations. After convergence, extract the joint probability distribution to reflect the global state estimate that fuses forward and backward inferences.

[0172] 5.3. Conduct multi-granularity state inference.

[0173] Extract state information from the constraint-adjusted probability distribution and calculate the uncertainty measure: For each state variable X, calculate the Shannon entropy H(X): H(X) = -∑_i P(X=xi)·log2(P(X=xi)). For the example order, the system calculates the probability distribution of the core state "Partial Liquidation Confirmed": P(Partial Liquidation Confirmed) = 0.82; P(Liquidation in Progress) = 0.12; P(Liquidation Failed) = 0.06; The corresponding entropy value H = -(0.82·log2(0.82) + 0.12·log2(0.12) + 0.06·log2(0.06)) = 0.76, indicating a medium level of uncertainty.

[0174] Generate state judgments according to different granularities: Business layer state: Determined as "Partial Liquidation Confirmed" (probability 0.82 > 0.7 threshold); Technical layer state: Partial Liquidation Confirmed (0.82), providing the complete probability distribution; Decision layer state: Order processing in progress, waiting for the remaining amount to be confirmed. The system finally generates the final state evaluation result: The order is currently in the "Partial Liquidation Confirmed" state, with a confidence level of 0.82, and the remaining 300 yuan needs additional verification, and it is expected to be completed within 10 - 15 minutes.

[0175] This embodiment successfully addresses the data delay and partial confirmation problems in the recharge order processing, showing the following advantages: Through the construction of a dynamic hypothesis space, the system successfully identifies and models the intermediate state outside the preset state machine of "Partial Liquidation Confirmed", solving the abnormal situations that cannot be expressed by the traditional fixed state space; By adopting the construction of a non-linear backtracking path, the system accurately locates the complex state transition path from "Payment Authorization Successful" to "Partial Liquidation Confirmed", avoiding the state judgment errors that may be caused by traditional linear backtracking; The differentiated backtracking strategy enables the system to prioritize the restoration of critical business fields (status flags and amounts), while delaying the backtracking of lower-priority data until the system resources are sufficient, ensuring business continuity and system efficiency; The forward-backward reasoning fusion enables the system to consider both the historical state path and the current observation evidence, accurately infer the true state of the order, and reasonably evaluate the expected time for the remaining processing; The multi-granularity state inference provides the appropriate form of state information for different levels of business decisions, including deterministic state judgments and uncertainty assessments, enhancing the usability and interpretability of the system.

[0176] The present invention completely solves the limitations of the fixed state space model by introducing a dynamic hypothesis space construction mechanism. The system can analyze the breakpoints in the order causal network in real time, calculate the state transition logic difference value matrix, and dynamically generate intermediate state nodes based on this. Combining with the non-standard states of similar cases in the historical anomaly pattern library, the system constructs an extended state set and applies a redundancy detection algorithm for optimization, finally forming a complete extended state network. This enables the system to break through the limitations of the preset state boundary, be able to flexibly handle various abnormal states and intermediate states, and improve its adaptability in complex business scenarios. The non-linear backtracking path construction method is adopted to overcome the limitations of traditional linear backtracking. The system applies an improved bidirectional breadth-first search algorithm, which is not limited to directly preceding nodes, but performs priority search based on the strength of causal relationships to identify potential state source points. In the path generation stage, the system allows the path to freely cross in the network, considering implicit dependencies and cross-branch relationships, and at the same time introduces a crossing cost coefficient to balance flexibility and optimization. Through a multi-dimensional filtering algorithm to evaluate the causal consistency, data support degree and historical frequency of the path, the system finally constructs a backtracking path graph with a complex topological structure, which can accurately handle complex scenarios such as state bifurcation, merging and cross-branch backtracking. By constructing a refined differential backtracking strategy execution mechanism, differential processing is implemented according to costs and business priorities. The system first calculates the comprehensive cost of each backtracking operation through a multi-factor cost evaluation model, considering dimensions such as computing resource consumption, business impact scope, state recovery difficulty and time sensitivity. Combining with the business priority stratification, the system constructs a 9×9 strategy mapping matrix to formulate personalized processing strategies for different types of backtracking operations, such as immediate full-scale backtracking for critical businesses, phased progressive backtracking for high-cost operations, etc. The system also applies a context-sensitive adjustment algorithm and a batch processing optimization algorithm to further refine the execution plan, ensuring to maximize the system resource utilization efficiency while guaranteeing the continuity of critical businesses. The inference fragmentation problem is solved through a forward-backward message passing mechanism. The system implements an improved Sum-Product algorithm on a bidirectional Bayesian network, designs a bidirectional message passing protocol, allowing evidence influence to be transmitted from past states to future states (forward) and from current observations to historical states (backward) simultaneously. For cyclic dependency structures, the system implements an improved Loopy Belief Propagation algorithm and introduces a message decay factor to effectively handle the problem of excessive information amplification in loop structures. When dealing with state bifurcation and merging points, the system applies special information distribution rules and weighted information fusion rules to integrate evidence according to the strength and reliability of causal associations. This enables the system to utilize both historical state paths and current observation evidence simultaneously, improving the accuracy and reliability of state evaluation.

[0177] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A method for real-time tracking of the status of a recharge order, characterized in that, It includes the following steps: Receive standardized order data, combine it with a pre-established order causal network model, generate an initial state evaluation result and abnormal index data, and construct a state transition hypothesis set based on this; Based on the order causal network model and the state transition hypothesis set, perform non-linear backtracking analysis to generate a state backtracking result; Read the state backtracking result, execute a differential backtracking strategy, generate an optimized backtracking plan and execute it to update the order state; Based on the state backtracking result and the updated order state, perform Bayesian network forward-backward inference fusion to generate a final state evaluation result; The steps of performing non-linear backtracking analysis to generate a state backtracking result include: When obtaining new delay data, integrate it with the standardized order data to generate updated state features, compare them with the initial state evaluation result, identify logical conflict points, and obtain a conflict identification result; Based on the conflict identification result and the order causal network model, construct a non-linear backtracking path to generate a backtracking path graph; Analyze the severity of the conflict identification result and the current system load status, determine the optimal backtracking depth, and generate a backtracking depth parameter; Verify the updated state features with each hypothesis path in the state transition hypothesis set, combine the backtracking path graph and the backtracking depth parameter to limit the verification range, obtain a hypothesis verification result and evaluate the impact, and select the optimal backtracking strategy based on this to generate a state backtracking result; The steps of constructing a non-linear backtracking path to generate a backtracking path graph include: Extract conflict point information from the conflict identification result, locate the corresponding state nodes in the order causal network model, construct a local network view, search for potential state source points in the historical direction along the conflict points from it, calculate the probability score of each node as a conflict source, and combine the state nodes exceeding a predetermined threshold to generate a preliminary backtracking point set; Based on the preliminary backtracking point set, construct possible backtracking paths from each backtracking point to the conflict point to generate a preliminary path set; Based on the preliminary path set, calculate the feasibility score for each path, combine the pre-stored path length, state transition probability, and historical similarity, generate a path set with scores, perform filtering and priority sorting, obtain qualified paths to form a directed graph, and obtain a backtracking path graph after adding metadata; The steps of generating a final state evaluation result include: Extract historical state evidence and real-time state evidence from the state backtracking result and the updated order state to generate a two-way evidence set; Convert the two-way evidence set into the probability distribution of nodes in a two-way Bayesian network to generate an evidence allocation result with credibility; at the same time, transmit from the past state to the future state and from the current observation to the historical state to generate a joint probability distribution; Adjust the joint probability distribution to satisfy the consistency rule constraints, generate a probability distribution after constraint adjustment, extract state information with different granularities from it to generate a multi-granularity state inference result, and combine the optimized backtracking plan and the latest observation evidence to generate a final state evaluation result.

2. The method according to claim 1, wherein The steps of constructing a state transition hypothesis set include: Analyze the abnormal index data and the initial state evaluation result, identify the breakpoint location result of the causal chain in the order causal network model, and perform historical pattern matching based on this to generate a historical reference path set; Combine the breakpoint location results to construct a dynamic hypothesis space and generate an extended state network; Based on the extended state network and the historical reference path set, perform multi-scale probabilistic hypothesis generation, obtain a preliminary state transition hypothesis set, optimize and prune it, and generate a final state transition hypothesis set.

3. The method according to claim 2, wherein The steps of constructing a dynamic hypothesis space and generating an extended state network include: Extract breakpoint information from the breakpoint location results, retrieve adjacent state nodes in the order causal network model, obtain a local network subgraph, and calculate the state transition logic difference value matrix between the breakpoints and adjacent nodes therein; According to the state transition logic difference value matrix, generate candidate intermediate states at positions where the difference value exceeds a predetermined threshold; Retrieve historical cases similar to the current breakpoint characteristics from the pre-stored historical abnormal pattern library and extract non-standard state nodes; Combine the candidate intermediate states and non-standard state nodes to construct a preliminary extended state set; For the preliminary extended state set, merge the extended states with feature similarity exceeding the preset threshold and perform causal consistency verification to generate an extended state network.

4. The method according to claim 3, characterized in that, The steps of calculating the state transition logic difference value matrix between the breakpoints and adjacent nodes in the local network subgraph include: Extract the multi-dimensional feature vectors of each node from the local network subgraph, For each pair of adjacent nodes, calculate the similarity of the multi-dimensional feature vectors to generate a preliminary similarity value; According to the state transition time series data between nodes, adjust the preliminary similarity value, increase the recent conversion weight, and generate a time series weighted similarity; Analyze the position and connection pattern of the nodes in the local network subgraph, adjust the calculation of the time series weighted similarity, and generate a structure fusion similarity; Convert the structure fusion similarity into a difference value, and use the method of subtracting the similarity from 1 to generate a state transition logic difference value matrix.

5. The method according to claim 1, characterized in that The steps of calculating the probability score of each node as a conflict source include: Based on the local network view and conflict point information, initialize the search queue, set the conflict point as the search starting point, and analyze the connection relationships in the local network view to assign conversion weights to each node, including forward and backward conversion weights, which represent the strengths of the cause and result respectively; Determine the optimal search depth according to the severity and type of the conflict point information; Extract similar conflict patterns from the historical conflict case library, calculate the frequency of each node as the source point in the historical similar conflicts, and generate a historical weight vector; Based on the conversion weights, the optimal search depth, and the historical weight vector, calculate the probability score of each node as a conflict source through an adaptive weighting formula.

6. The method according to claim 1, wherein The steps of executing a differential backtracking strategy to generate an optimized backtracking plan include: Read the state backtracking results, calculate the backtracking cost of each backtracking operation, and generate a backtracking cost matrix; Analyze the business scenarios and user experience impacts involved in the state backtracking results, combine the predefined business priority rules, classify the backtracking operations by priority, and generate business priority tags; formulate differential processing strategies for different types of backtracking operations in combination with the backtracking cost matrix to generate a differential backtracking strategy table; According to the differential backtracking strategy table and the current system load status, optimize the execution order and resource allocation of the backtracking operations to generate a resource scheduling plan; Generate an optimized backtracking plan by integrating the differential backtracking strategy table, resource scheduling plan, and business priority tags.

7. The method according to claim 6, wherein The steps to generate the differential backtracking strategy table include: Map the backtracking operations in the backtracking cost matrix to the range of 0 to 1 according to the cost value, and divide them into three cost levels; Divide the operations in the business priority tags into three priority levels according to the business importance; Construct a strategy mapping matrix with the cost levels as rows and the priority levels as columns; According to the position of each backtracking operation in the strategy mapping matrix, assign a preliminary strategy type, and for each preliminary strategy type, refine the specific execution parameters to obtain a preliminary strategy; Perform fine-grained adjustment on the preliminary strategy and merge similar backtracking operations to obtain an adjusted strategy; Based on the adjusted strategy, generate an execution specification for each backtracking operation, and integrate them to obtain the differential backtracking strategy table.

8. The method according to claim 7, wherein The steps to construct the strategy mapping matrix include: Extract the basic backtracking strategy template from the system preset strategy library; Perform fine-grained division on the cost levels to obtain fine cost levels; Apply context-sensitive enhancement to the priority levels according to the current business scenario data to generate enhanced priority levels; Based on the fine cost levels, enhanced priority levels, and basic backtracking strategy template, construct an initial strategy mapping matrix; Detect the mutation points between adjacent strategies in the initial strategy mapping matrix, identify strategy conflicts, generate a conflict resolution matrix and perform smoothing processing to generate the strategy mapping matrix.

Images