Cross-platform business process automatic bridging method for heterogeneous enterprise system

Abstract

According to the method, the service node event of the first service platform is monitored in real time, the current state information, the metadata and the change history of the second service platform are synchronously obtained, and the total state context of the first service platform is collected; constructing the double-platform state features into a time sequence feature tensor, inputting the time sequence feature tensor into a Transform sequence generation model, analyzing cross-platform semantic association through a hierarchical attention mechanism, and generating an optimal atomic service instruction sequence based on multi-task learning; inputting the instruction sequence into a graph neural network discrimination model, predicting potential conflicts and compliance risks through a dynamic service graph and risk analysis, and outputting a risk score and an optimization scheme; and dynamically adjusting an execution strategy according to the risk score, constructing triple feedback data, and continuously optimizing model parameters through online learning. According to the method, the problem of low collaboration efficiency caused by state synchronization delay and semantic association complexity among heterogeneous systems is solved, and high efficiency, accuracy and safety of business process bridging are realized.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method for automating cross-platform business process bridging for heterogeneous enterprise systems. Background Technology

[0002] Multi-platform business collaboration is a crucial area in modern digital transformation, directly impacting the efficiency and stability of cross-system collaboration within enterprises. In industries such as finance, logistics, and e-commerce, different business platforms require real-time interaction to complete complex tasks, such as order processing, payment settlement, or supply chain management. Efficient collaboration mechanisms can significantly improve operational efficiency and reduce costs caused by information asymmetry or operational errors. However, current multi-platform collaboration solutions often fall short in handling complex business scenarios due to technological limitations, failing to meet the dual requirements of real-time performance and accuracy. Existing methods face significant limitations in cross-platform collaboration. Many solutions rely on static rules or preset templates to generate instructions, making it difficult to adapt to dynamically changing business environments. For example, in cross-border e-commerce, inventory updates between platforms may be out of sync due to network latency, leading to incorrect order allocation. Furthermore, existing methods often lack a comprehensive assessment of potential risks; for instance, in financial transactions, failure to identify compliance issues before execution can easily lead to financial losses or legal risks. These problems stem from insufficient dynamic correlation analysis of cross-platform states, limiting the improvement of collaboration efficiency. The core technical challenge lies in how to accurately analyze the semantic relationships between cross-platform states. Semantic relationships refer to the logical relationships between the states of business objects (such as inventory, order status, and account balance) across different platforms. For example, in an e-commerce scenario, after a buyer places an order, the payment platform needs to confirm the fund status, and the logistics platform needs to check the inventory and allocate transportation resources. If the semantic relationships between these states are not accurately captured, it may lead to incorrect instruction generation, such as assigning out-of-stock items to customers. This problem stems from the temporal and complex nature of states between platforms: state data not only changes rapidly over time but also involves logical dependencies between multiple parties, making it difficult to parse in real time using traditional methods. Therefore, the issues of state synchronization and semantic analysis in business processes are particularly prominent. Taking financial payments as an example, cross-border transfers need to ensure that the initiator's account balance is sufficient and the recipient's compliance is correct, while also considering exchange rate fluctuations and transaction timeliness. If the balance information is incorrect due to state synchronization delays, or compliance requirements are ignored due to insufficient semantic relationship analysis, the transfer may fail or lead to legal disputes. How to capture and accurately parse the semantic relationships between cross-platform states in real time in a dynamic and ever-changing business environment has become a key issue in improving the efficiency and security of multi-platform business collaboration. Summary of the Invention

[0003] This invention provides a method for automating cross-platform business process bridging for heterogeneous enterprise systems, mainly including: Multi-source context awareness step: Real-time monitoring of predefined business node events in the first business platform; in response to event triggering, synchronously acquiring the current state information, metadata, and change history of the corresponding business object in the second business platform through standardized interfaces, while simultaneously collecting the full state context of the corresponding business object in the first business platform; Intelligent instruction generation step: Constructing a temporally sequenced feature tensor from the multi-dimensional state features from both platforms, and inputting it into a sequence-to-sequence generation model based on the Transformer architecture; This model analyzes the semantic relationships between cross-platform states through an attention mechanism, and generates an atomic business instruction sequence that best matches the current business context based on multi-task learning of historical successful bridging patterns; Dynamic risk verification step: Inputting the generated instruction sequence into a discriminative model based on graph neural networks for multi-dimensional verification; This model predicts potential conflicts and compliance risks in the instruction execution path by modeling the topological relationships between business objects, and outputs a risk score and optimization scheme; Adaptive execution and evolution step: Dynamically adjusting the instruction execution strategy according to the risk score, executing the verified instruction sequence, and constructing a triplet feedback data containing the instruction sequence, execution environment snapshot, and performance indicators; Continuously optimizing the parameter weights of the generation model and the discriminative model through an online learning mechanism. Furthermore, the intelligent instruction generation step includes the following sub-steps: Feature tensor construction sub-step: Normalize the state information from the two platforms, map the classification features into dense vectors through the embedding layer, perform adaptive bucketing normalization on the numerical features, extract periodic pattern representations from the temporal features through a long short-term memory network, and finally concatenate them to form a temporal feature tensor; Hierarchical attention calculation sub-step: The generation model first analyzes the temporal dependency of state changes within a single business platform using a causal masking mechanism through the intra-platform attention layer, then calculates the correlation weight between the state features of the first business platform and the state features of the second business platform using a cross-platform attention mechanism through a cross-platform attention layer, and finally dynamically fuses the output results of intra-platform attention and cross-platform attention through a gating mechanism; Multi-task instruction generation sub-step: The generation model synchronously executes the main task to generate atomic business instruction sequences, while simultaneously predicting the state transition of the business platform at the next moment through an auxiliary task.Furthermore, the dynamic risk verification step includes the following sub-steps: Business graph construction sub-step: using business objects as nodes and business relationships between objects as edges, a dynamic business graph with temporal weights is constructed; Risk propagation analysis sub-step: the graph neural network adopts a temporal graph attention network architecture, and performs multi-hop information aggregation along the business graph through a message passing mechanism, prioritizing the analysis of business objects on critical paths dynamically identified by the PageRank algorithm, and calculating the cascading effects that instruction execution may cause; Multi-dimensional risk assessment sub-step: for detected high-risk instruction sequences, alternative path suggestions are automatically generated as optimization solutions, and the business compliance, resource conflict probability, and performance bottleneck risk of the instruction sequence are comprehensively evaluated to generate an overall risk score. Furthermore, the adaptive execution and evolution steps include the following sub-steps: Execution strategy selection sub-step: Based on risk scores, the instruction sequence is divided into three levels: high, medium, and low risk, corresponding to three strategies: execution after manual review, restricted automatic execution, and fully automatic execution, respectively; Feedback data construction sub-step: Recording system state snapshots before and after instruction sequence execution, collecting performance data on execution time, resource consumption, and changes in business metrics, and constructing triplet feedback samples; Incremental learning optimization sub-step: Using a sliding window mechanism to select recent feedback data, setting model update triggers to initiate model optimization when the time window set to balance model real-time performance and stability, the sample quantity threshold set to ensure sufficient training data, or the model performance degradation exceeds a preset threshold, employing an elastic weight consolidation algorithm to prevent catastrophic forgetting, and coordinating model updates across multiple business units through a federated learning architecture. Furthermore, it also includes an elastic degradation step: continuously monitoring the service status and output confidence of the generated model; when an abnormal model service or a confidence level consistently below a threshold is detected, automatically switching to a rule engine-based backup instruction generation mode; environmental data collected during degradation is quickly integrated into the training process after model recovery through a priority replay mechanism.This invention provides a cross-platform business process automation bridging system for heterogeneous enterprise systems, comprising: a multi-source context awareness module, configured to monitor predefined business node events in a first business platform in real time, and in response to event triggering, synchronously acquire the current state information, metadata, and change history of the corresponding business object in a second business platform through a standardized interface, while simultaneously collecting the full state context of the corresponding business object in the first business platform; and an intelligent instruction generation module, configured to construct temporally sequenced feature tensors from multi-dimensional state features from both platforms, input them into a sequence-to-sequence generation model based on the Transformer architecture, and analyze the semantic relationships between cross-platform states through an attention mechanism. The invention utilizes multi-task learning of historical successful bridging patterns to generate atomic business instruction sequences optimally matched to the current business context. A dynamic risk verification module is configured to input the generated instruction sequences into a graph neural network-based discriminative model for multi-dimensional verification. By modeling the topological relationships between business objects, it predicts potential conflicts and compliance risks in the instruction execution path and outputs risk scores and optimization schemes. An adaptive execution and evolution module is configured to dynamically adjust the instruction execution strategy based on the risk score, execute verified instruction sequences, and construct triplet feedback data containing the instruction sequences, execution environment snapshots, and performance indicators. It continuously optimizes the parameter weights of the generation and discriminative models through an online learning mechanism. This invention provides an electronic device, including a processor, a memory, and a computer program stored in the memory, characterized in that the processor implements the above method when executing the program.

[0004] The present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that the program implements the above-described method when executed by a processor.

[0005] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses an intelligent instruction generation and execution system for cross-platform business collaboration. Addressing the low collaboration efficiency caused by state synchronization delays, complex semantic relationships, and potential execution risks across multiple business platforms, it seamlessly integrates four modules: multi-source context awareness, intelligent instruction generation, dynamic risk verification, and adaptive execution and evolution, constructing a logically coherent solution. First, it monitors and synchronizes the states and historical data of business objects on both platforms in real time, constructing a temporal feature tensor. This tensor is input into a Transformer-based sequence-to-sequence model, and a hierarchical attention mechanism is used to accurately analyze cross-platform state semantic relationships, generating the optimal atomic instruction sequence. Subsequently, dynamic risk verification based on graph neural networks predicts potential conflicts and compliance risks through business graph modeling and multi-hop information aggregation, outputting risk scores and optimization schemes. Finally, it dynamically adjusts the execution strategy and continuously optimizes the model through online learning, ensuring efficient and stable operation. This invention, through the deep integration of cross-platform semantic analysis and risk prediction, achieves high efficiency, accuracy, and security in instruction generation, significantly improving the efficiency and robustness of multi-platform business collaboration. Attached Figure Description

[0006] Figure 1 This is a flowchart of a cross-platform business process automation bridging method for heterogeneous enterprise systems according to the present invention. Detailed Implementation To further understand the content of this invention, a detailed description of the invention is provided in conjunction with the accompanying drawings and embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0007] like Figure 1 This embodiment provides a method for automating cross-platform business process bridging for heterogeneous enterprise systems, including: S1. Multi-source context awareness steps: Real-time monitoring of predefined business node events in the first business platform; in response to event triggering, synchronous acquisition of the current status information, metadata and change history of the corresponding business object in the second business platform through standardized interfaces; and collection of the full state context of the corresponding business object in the first business platform.

[0008] In one embodiment, step S11 defines business node events as preset trigger points such as order submission or inventory update, and continuously scans the log stream or API endpoint of the first business platform through an event listener to capture the occurrence time of these events.

[0009] Specifically, step S111 involves activating a response mechanism based on the captured event type. For example, when an order submission event occurs, the business object identifier associated with the event is immediately queried.

[0010] Step S112: Using standardized interfaces such as RESTful API or GraphQL, send a synchronization request to the second business platform to obtain the current status information of the corresponding business object, such as inventory level or processing stage, metadata such as object tags or permission settings, and change history, which includes timestamps and a sequence of modification details.

[0011] Step S113: Simultaneously collect the full state context within the first business platform, covering the attribute values, relationships, and environment variables of business objects, such as user roles or system load.

[0012] For example, this step ensures cross-platform data consistency through real-time monitoring, which helps reduce latency and improve bridging efficiency.

[0013] In one possible implementation, considering an enterprise supply chain scenario, when the first platform detects a supplier order event, the interface synchronously pulls the material status from the second platform and collects the order context from the first platform to form a complete view and avoid information silos.

[0014] In one embodiment, in step S11, the event listener employs the Kafka streaming framework to continuously scan logs to capture events. Here, Kafka refers to a distributed event streaming platform that consumes business node event data streams in real time through a topic subscription mechanism, ensuring high throughput and low latency monitoring. When activating a response in step S111, an event-associated identifier, such as a unique order ID, is used to locate the object. The standardized interface request process in step S112 involves constructing a JSON payload containing query parameters, and then parsing the status information, metadata, and change history sequence in the response. For example, the change history is represented by a time-ordered list of each modified field and value. The full state context collected in step S113 is obtained through a database snapshot or memory cache, forming a structured data packet.

[0015] Specifically, this implementation method is beneficial for data synchronization between heterogeneous systems. In inventory management applications, when an event is triggered, quickly obtaining the change history helps track supply chain changes and optimize the generation of subsequent instructions.

[0016] For example, assuming the first platform is an ERP system and the second platform is a CRM system, when an order node event is triggered in the ERP system, the interface obtains the status of the customer object in the CRM system, such as activity level, metadata such as contact information tags, and change history such as recent interaction records. Simultaneously, it collects the full context of the order in the ERP system, such as amount and item list. This multi-source awareness improves business continuity and reduces manual intervention.

[0017] In one embodiment, for high-load scenarios, step S11 is expanded to listen to events from multiple nodes in parallel, managing the scanning tasks through a thread pool. Step S111 prioritizes responses according to event priority, for example, prioritizing urgent inventory events. Step S112 interface synchronization includes a timeout retry mechanism to ensure the acquisition of a complete change history. Step S113 context acquisition adds encryption processing to protect sensitive data.

[0018] Specifically, this extension benefits large-scale enterprise environments by providing robustness. In the case of supply chain disruptions, the sensing step captures data and synchronizes it after the event, helping to respond quickly to market fluctuations.

[0019] S2. Intelligent Instruction Generation Steps: Multi-dimensional state features from both platforms are constructed into temporally sequenced feature tensors and input into a sequence-to-sequence generation model based on the Transformer architecture. This model analyzes the semantic relationships between cross-platform states through an attention mechanism and, based on multi-task learning of historical successful bridging patterns, generates an atomic business instruction sequence optimally matched to the current business context. The atomic business instruction referred to in this paper is the smallest business operation unit that can be directly executed in the target business platform. This operation possesses transaction atomicity, meaning its execution result has only two states: success or failure, without generating a partially successful intermediate state. An atomic business instruction typically corresponds to an atomic API call or a database transaction operation provided by the target platform. For example, in an inventory management system, "decrease the inventory quantity of the product with SKU XXX by 1" is an atomic business instruction; while in an order system, "update the order status to 'paid'" is another atomic business instruction. The instruction sequence generated by this invention is composed of multiple such atomic business instructions combined in a logical order.

[0020] Step S21: Normalize the state information from the two platforms, map the classification features into dense vectors through the embedding layer, perform adaptive bucketing normalization on the numerical features, and extract the periodic pattern representation of the temporal features through the long short-term memory network. Finally, the temporal features are concatenated to form a temporal feature tensor.

[0021] In one embodiment, when normalizing the status information in step S21, event trigger data from the first business platform and the corresponding object status from the second business platform are collected first. For example, in a supply chain system, the first platform records inventory changes, and the second platform records order status. Classification features, such as business type, are converted into fixed-dimensional vector representations through an embedding layer to ensure comparability of category data across different platforms.

[0022] Step S211: For numerical features such as inventory quantity or order amount, use an adaptive binning method to dynamically divide the range according to the data distribution. For example, divide the inventory quantity into three bins: low, medium and high, and normalize them to the range of 0 to 1.

[0023] Step S212: For time-series features such as state change history, input the data into a long short-term memory network. This network retains long-term dependencies through forget gates, input gates, and output gates, and extracts periodic patterns such as daily order peaks.

[0024] For example, in enterprise resource planning systems, this construction method ensures that the feature tensor captures cross-platform temporal dynamics, which is beneficial to the stability of subsequent model inputs.

[0025] In step S22, the generative model first uses a causal masking mechanism to analyze the temporal dependency of state changes within a single business platform through an intra-platform attention layer, then uses a cross-attention mechanism to calculate the correlation weight between the state features of the first business platform and the state features of the second business platform through a cross-platform attention layer, and finally uses a gating mechanism to dynamically fuse the output results of intra-platform attention and cross-platform attention.

[0026] In one embodiment, the platform-in-attention layer in step S22 uses a causal mask to allow the current position to focus on previous positions, for example, when analyzing the inventory time series of the first platform, to ensure that future data does not affect the current calculation.

[0027] In step S221, in the cross-platform attention layer, cross-attention calculation is performed to query features from the first platform, with keys and values ​​from the second platform, generating association weights, such as the matching degree between supply chain orders and financial payments.

[0028] Step S222: A gating mechanism, such as a sigmoid function, is used to weight and fuse the two attention outputs, adjusting the weights to highlight key correlations.

[0029] For example, in bridging finance and human resources systems, this integration improves the accuracy of instruction generation and reduces errors caused by cross-platform inconsistencies.

[0030] In step S23, the generation model synchronously executes the main task to generate atomic business instruction sequences, and at the same time predicts the state transition of the business platform at the next moment through auxiliary tasks.

[0031] In one embodiment, in step S23, the main task outputs instructions such as "update inventory records" and the auxiliary task predicts status such as "order completion probability".

[0032] For example, this multi-task learning, trained on historical bridging data, improves the model's predictability of future states and is beneficial to the robustness of bridging.

[0033] In one embodiment, for heterogeneous systems such as a manufacturing execution system and a customer relationship management system, when step S2 is applied as a whole, a feature tensor is first constructed to capture the temporal sequence of production status and customer feedback. Then, a Transformer model analyzes the correlation and generates instruction sequences such as automated approval processes. This approach reduces latency in high-load scenarios, such as when processing thousands of orders daily. The attention mechanism prioritizes identifying critical paths to ensure that instructions match the context.

[0034] For example, in supply chain bridging, after the model learns historical patterns, the generated instructions avoid manual intervention and optimize response time.

[0035] In one embodiment, for parameter tuning, the number of Transformer layers is set to 6, the embedding dimension is 512, and the multi-task loss function combines cross-entropy and mean squared error. During training, historical successful bridging data, such as 1000 bridging records from the past year, is used. This setup, in a financial auditing scenario, ensures compliance of instruction sequences and reduces risk.

[0036] For example, when analyzing semantic relationships using attention mechanisms, calculating weight matrices reveals inter-platform dependencies, such as how payment delays affect inventory updates, and generating optimized instruction sequences improves efficiency.

[0037] In one embodiment, during multi-task learning, the main task generates instructions such as "synchronize data update," and the auxiliary task is triggered to execute when the predicted state transition probability is above 0.8. This mechanism is beneficial for seamlessly bridging heterogeneous data sources in system integration after corporate mergers and acquisitions.

[0038] S3. Dynamic Risk Verification Step: The generated instruction sequence is input into a discriminative model based on graph neural networks for multi-dimensional verification. This model predicts potential conflicts and compliance risks in the instruction execution path by modeling the topological relationships between business objects, and outputs risk scores and optimization solutions.

[0039] Specifically, this step first constructs a relationship graph between business objects, then performs risk propagation calculations, and finally evaluates and generates outputs to ensure the secure execution of instruction sequences in cross-platform business processes.

[0040] Step S31 involves constructing a dynamic business graph with temporal weights, using business objects as nodes and business relationships between objects as edges. In this sub-step, business objects from the first and second business platforms are represented as graph nodes, such as order processing objects or inventory management objects. Each node contains current status information, such as inventory level or order status. Business relationships, such as orders depending on inventory, are represented as edges, and each edge is assigned a temporal weight based on change history; for example, edges with more recent changes have higher weights to reflect dynamic changes. Through this construction, the graph captures the real-time topology of business objects, forming the basis for subsequent risk prediction. This graph construction method helps identify potential cascading risks and improves the robustness of the bridging method.

[0041] In one embodiment, the construction process of step S31 is further refined. Step S311 involves collecting the current state information and change history of business objects on both platforms, such as obtaining the update timestamp and status value of order objects from a standardized interface. Step S312 involves calculating temporal weights, for example, using an exponential decay function to weight the change history, with recent changes receiving a weight of 0.8 and earlier changes receiving 0.2, ensuring the graph reflects the latest dynamics. Step S313 involves assembling the graph, connecting nodes and weighted edges into a dynamic structure. This refinement allows the graph to more accurately simulate business interactions in heterogeneous systems, improving the accuracy of subsequent analysis.

[0042] For example, in an enterprise supply chain scenario, assuming the first platform is a procurement system and the second platform is an inventory system, when constructing the graph in step S31, purchase orders are used as nodes, available inventory is used as another node, edges represent dependencies, and the time-series weight is calculated as 0.9 based on the most recent inventory update time. This construction ensures the timeliness of risk verification, avoids misjudgments caused by outdated data, and thus improves the overall bridging efficiency.

[0043] Step S32: The graph neural network adopts a temporal graph attention network architecture. Through a message passing mechanism, it performs multi-hop information aggregation along the business graph, prioritizing the analysis of business objects on key paths dynamically identified by the PageRank algorithm, and calculating the cascading effects that instruction execution may trigger. The temporal graph attention network is a variant of the graph neural network that incorporates a time dimension when processing graph data, allocating importance weights between nodes through an attention mechanism. The specific process includes initializing node embeddings and converting the state of business objects into vector representations; then performing message passing, aggregating information from neighboring nodes at each hop (e.g., the first hop aggregates the states of directly connected nodes, and the second hop extends to indirect associations); the PageRank algorithm is used to calculate node importance scores, identifying key paths by iteratively calculating the in-degree link probability of nodes, for example, paths with scores higher than 0.5 are prioritized; finally, cascading effects are calculated, such as how an instruction modifying inventory may have a cascading impact on downstream orders. In this way, the model predicts potential conflicts, such as resource contention or state inconsistencies.

[0044] In one embodiment, the risk propagation analysis in step S32 is extended to different aggregation hop counts. Step S321 involves applying the PageRank algorithm to identify critical paths, for example, prioritizing order-inventory links with a path score of 0.6. Step S322 involves multi-hop aggregation, where the first hop aggregates direct neighbors using an attention weight of 0.7, and the second hop expands with a weight of 0.3. Step S323 quantifies cascading effects; for example, if a sequence of instructions causes inventory to fall into negative territory, it is marked as high impact. This extension adapts to complex enterprise systems, improves the comprehensiveness of forecasts, and helps reduce unexpected disruptions during execution.

[0045] For example, in financial transaction bridging, when processing the instruction sequence in step S32, PageRank identifies the transaction confirmation path as critical. Message passing aggregation shows that if a transfer instruction conflicts with an existing balance, the cascading impact score reaches 0.8, leading to potential compliance risks. This analysis ensures early risk identification and improves system stability.

[0046] Step S33: For detected high-risk instruction sequences, automatically generate alternative path suggestions as optimization solutions, and comprehensively evaluate the scores of the instruction sequences across three dimensions: business compliance, resource conflict probability, and performance bottleneck risk, to generate an overall risk score. In this sub-step, high-risk sequences refer to those whose cascading effects exceed a threshold, such as 0.7; alternative paths are generated by searching for spare edges in the graph, such as bypassing conflict nodes; evaluation dimensions include checking for compliance violations of business rules, calculating the probability of resource overlap conflicts, and estimating execution time bottlenecks; the final score is a weighted average of the three dimensions, such as a compliance weight of 0.4. This evaluation output guides subsequent execution adjustments.

[0047] For example, in a manufacturing enterprise system, step S33 evaluates a production scheduling instruction sequence and finds a resource conflict probability of 0.6. It then generates alternative path suggestions to adjust the scheduling order, reducing the overall risk score to 0.3. This method optimizes the bridging process and reduces actual risk.

[0048] S4. Adaptive Execution and Evolutionary Steps: Dynamically adjust the instruction execution strategy based on the risk score, execute the verified instruction sequence, and construct a triplet feedback data containing the instruction sequence, execution environment snapshot, and performance indicators. Continuously optimize the parameter weights of the generation model and the discriminant model through an online learning mechanism.

[0049] In one embodiment, the adaptive execution and evolution steps include the following sub-steps.

[0050] S41 Execution Strategy Selection Sub-step: Based on the risk score, the instruction sequence is divided into three levels: high, medium, and low risk, corresponding to three strategies: execution after manual review, restricted automatic execution, and fully automatic execution, respectively.

[0051] Specifically, the risk score is derived from the output of the aforementioned dynamic risk verification steps. For example, a risk score exceeding 0.8 is classified as high risk, requiring manual review to ensure compliance; a score between 0.4 and 0.8 is classified as medium risk, employing restricted automatic execution, such as setting real-time monitoring points during execution to allow for interruption; and a score below 0.4 is classified as low risk, allowing for fully automatic execution. This classification helps balance automation efficiency with risk control.

[0052] S411 Collect and classify risk scores: Obtain the overall risk score from the output of the graph neural network and classify it into levels according to preset thresholds.

[0053] S412 is mapped to the execution strategy: for high-risk levels, a notification mechanism is triggered to send the notification to the administrator for review; for medium-risk levels, resource restrictions are added before execution, such as timed execution windows; for low-risk levels, the code is directly injected into the execution queue.

[0054] For example, in the bridging scenario between an enterprise ERP system and a CRM system, high-risk instruction sequences may involve the synchronization of financial data. If the score is high, manual review is required to prevent data leakage, thereby improving system security.

[0055] S42 Feedback Data Construction Sub-step: Record system state snapshots before and after instruction sequence execution, collect performance data such as execution time, resource consumption, and changes in business indicators, and construct triplet feedback samples.

[0056] It should be noted that the specific construction process of the triplet feedback samples includes storing the instruction sequence as the first element, the execution environment snapshot (such as CPU utilization and memory usage) as the second element, and the performance indicators (such as response time and success rate) as the third element in a serialized manner as feedback data, so as to ensure that subsequent optimization steps can directly utilize these samples.

[0057] S421 records snapshots before and after execution: captures the platform state before instruction execution, including the current value of business objects and network latency, and then immediately captures the changed state after execution.

[0058] S422 collects performance data: measures the total time from command initiation to completion, peak resource consumption, and changes in business metrics such as transaction completion rate.

[0059] S423 Assemble Triples: Integrate the above data into structured triples and add timestamps to support time series analysis.

[0060] In one embodiment, for supply chain management bridging, if the business metrics show a 5% increase in transaction completion rate after the instruction sequence is executed, the triplet feedback sample records this change, which is beneficial for the model to learn successful patterns.

[0061] S43 Incremental Learning Optimization Sub-step: A sliding window mechanism is used to select recent feedback data. Model update triggers are set to start model optimization when the time window set to balance the real-time performance and stability of the model, the sample number threshold set to ensure sufficient training data, or the model performance decay exceeds a preset threshold. An elastic weight consolidation algorithm is used to prevent catastrophic forgetting, and model updates of multiple business units are coordinated through a federated learning architecture.

[0062] Specifically, the sliding window mechanism refers to maintaining a fixed-size time window, such as feedback data from the most recent 24 hours, to select recent triplet feedback samples; model update triggers include time windows (e.g., checking every 6 hours), sample number thresholds (e.g., accumulating 100 samples), or performance degradation thresholds (e.g., triggering when accuracy drops by more than 10%). The elastic weight consolidation algorithm is a method to prevent catastrophic forgetting by assigning higher weights to older knowledge during optimization to preserve previously learned parameters; the federated learning architecture allows different business units to update model parameters locally and then aggregate the global model without sharing the original data.

[0063] S431 Select Feedback Data: Use a sliding window to extract the most recent triplet feedback sample from storage to ensure data freshness.

[0064] S432 checks the trigger conditions: whether the monitoring time window has expired, whether the number of samples has reached the threshold, or whether the performance indicators have degraded. If any one of these conditions is met, optimization will be initiated.

[0065] S433 performs parameter optimization: using selected triplet feedback samples as input, the Transformer parameters of the generator model and the graph neural network weights of the discriminant model are updated through gradient descent. At the same time, the elastic weight consolidation algorithm is applied to calculate the weighted average of the old parameters to prevent forgetting.

[0066] S434 Coordinated Federated Update: Each business unit uploads local parameter changes, and the central node aggregates and distributes the updated model.

[0067] For example, in financial transaction bridging, if the model performance degrades beyond a threshold, optimization is triggered by using triplet feedback samples to adjust parameters, which helps improve the accuracy of instruction generation and reduce cross-platform conflicts.

[0068] In one embodiment, for the trigger of the incremental learning optimization sub-step, when the sample number threshold is set to 50, if enough triplet feedback samples are accumulated, the optimization process is started, and the model weights are fine-tuned using these samples, thereby achieving real-time adaptation to business changes.

[0069] It should be noted that this setup is beneficial in data-intensive scenarios such as e-commerce order synchronization, enabling rapid response to changes in patterns and preventing models from becoming outdated.

[0070] For example, in bridging the manufacturing execution system and the inventory system, if performance degrades within a time window, the optimized model can better predict risks, improving overall bridging efficiency by 15%.

[0071] In one embodiment, when applying the elastic weight consolidation algorithm, the feedback samples of historically successfully bridged triples are assigned a weight of 1.5 times during optimization to ensure that new learning does not overwrite old knowledge, which is beneficial to long-term stability.

[0072] It should be noted that this algorithm is based on the gradient projection principle, and during the calculation, the new gradient and the old parameters are projected onto a compatible subspace.

[0073] For example, in HR management system bridging, this approach prevents the forgetting of early learning patterns and maintains cross-platform consistency.

[0074] S5 also includes an elastic degradation step: continuously monitor the service status and output confidence of the generated model; when an abnormal model service is detected or the confidence level is continuously lower than the threshold, automatically switch to the backup instruction generation mode based on the rule engine; environmental data collected during degradation is quickly integrated into the training process through a priority replay mechanism after the model recovers.

[0075] In one embodiment, the resilient degradation step in step S5 specifically includes the following sub-steps.

[0076] Step S51: Continuously monitor the service status and output confidence of the generated model.

[0077] Specifically, by deploying a monitoring agent in the runtime environment of the generative model, the model's service status indicators such as CPU utilization, memory usage, and response latency are checked every preset time interval, such as 10 seconds. At the same time, a confidence score is attached to each atomic business instruction sequence generated by the model. This score is calculated based on the output vector of the self-attention layer of the Transformer architecture. That is, after summing the attention weights through the softmax function and normalizing, a value between 0 and 1 is obtained, which represents the reliability of the generated result.

[0078] In one embodiment, the monitoring process also includes a logging mechanism that writes service status metrics and confidence scores into a distributed log system in real time to facilitate subsequent anomaly diagnosis.

[0079] Step S52: When an abnormality is detected in the model service or the confidence level is continuously lower than the threshold, automatically switch to the backup instruction generation mode based on the rule engine.

[0080] Specifically, model service anomalies are defined as CPU utilization exceeding 90% for 30 seconds or response latency greater than 500 milliseconds. The confidence threshold is set to 0.7. If the confidence of three consecutive generated results is lower than this value, the switching logic is triggered. During the switching, the system first pauses the input stream of the generated model and redirects the current multidimensional state feature tensor to the rule engine. This engine predefines a series of if-then rules. For example, if the business node event of the first business platform is "order creation" and the corresponding object status of the second platform is "sufficient inventory", then the generated instruction sequence includes "synchronize inventory deduction" and "notify shipment".

[0081] For example, in cross-platform business automation scenarios, if the generated model causes service anomalies due to network fluctuations, the system will automatically activate the rule engine after detecting that the response delay exceeds the threshold. It will then use preset rules to match the current feature tensor and generate a backup instruction sequence to ensure that the business process is not interrupted. This will improve the robustness of the system and reduce business losses caused by downtime.

[0082] In one embodiment, the switching process also involves a smooth transition mechanism, namely, buffering the most recent feature tensors before switching and prioritizing the processing of these buffered data in the rule engine mode to maintain the continuity of instruction generation.

[0083] Step S521: Determine the anomaly type based on the monitoring results. If the service status is abnormal, immediately isolate and generate a model instance and restart the standby instance. If the confidence level is abnormal, accumulate the anomaly count and perform a switchover when the count reaches 5.

[0084] Step S522: When activating the rule engine, extract the rule template from the historical successful bridging mode and dynamically fill the parameters according to the current time series feature tensor. For example, adaptively bucket the numerical features in the feature tensor and map them to the rule conditions.

[0085] For example, when the rule engine is processing order synchronization between heterogeneous enterprise systems, if it detects that the confidence level is lower than 0.7, it will switch to using a rule such as "if the change history shows that the inventory change exceeds 10%, then add a verification instruction". This can quickly generate reliable instructions, avoid business errors caused by low confidence output of the model, and improve the overall system stability through this backup.

[0086] Step S53: Environmental data collected during the degradation period is quickly incorporated into the training process through a priority replay mechanism after the model is restored.

[0087] Specifically, the environmental data during the degradation period includes the state context collected by the multi-source context-aware steps, the instruction sequence generated by the rule engine and its execution results, such as execution time and changes in business indicators; model recovery is defined as starting priority replay after the service state indicators return to normal and the confidence level exceeds 0.8, that is, using a priority queue to sort the collected data, with high-priority data being those samples that deviate significantly from the historical successful bridging patterns. These data are first converted into triplet feedback samples, and the parameter weights of the generated model are updated through online learning mechanisms such as gradient descent.

[0088] For example, in business process automation bridging, high-risk instruction execution data collected during the degradation period can be replayed first after the model recovers. This can quickly optimize the model's adaptability to abnormal scenarios, reduce the frequency of future degradations, and thus improve the long-term reliability of cross-platform systems.

[0089] In one embodiment, the replay mechanism also includes batch processing optimization, which integrates high-priority environmental data into training in batches to avoid the computational overhead caused by loading all at once.

[0090] Step S531: When collecting environmental data during the degradation period, construct an extended triplet containing state context, backup instruction sequence and performance indicators, and label it with priority labels based on the deviation degree, i.e. the Euclidean distance of the difference from the standard generation model output.

[0091] Step S532: After the model is restored, high-priority extended triples are popped out using a priority queue and input into the multi-task learning framework to update the sequence-to-sequence generation model parameters of the Transformer. At the same time, to prevent catastrophic forgetting, the original weights are retained by adding a regularization term.

[0092] For example, suppose that during the downgrade, the environmental data of an inventory synchronization event is processed. The high bias is due to the rule engine generating an instruction sequence that has an extra step of verification compared to the model's expectations. During playback, the model's cross-platform attention layer is updated first, which can improve the semantic association analysis of similar events and bring more accurate instruction generation results.

[0093] In one embodiment, this priority playback can also be linked with other steps, such as sharing data with the dynamic risk verification module to further optimize the discriminative model of the graph neural network.

[0094] This invention provides a cross-platform business process automation bridging system for heterogeneous enterprise systems, comprising: a multi-source context awareness module, configured to monitor predefined business node events in a first business platform in real time, and in response to event triggering, synchronously acquire the current state information, metadata, and change history of the corresponding business object in a second business platform through a standardized interface, while simultaneously collecting the full state context of the corresponding business object in the first business platform; and an intelligent instruction generation module, configured to construct temporally sequenced feature tensors from multi-dimensional state features from both platforms, input them into a sequence-to-sequence generation model based on the Transformer architecture, and analyze the semantic relationships between cross-platform states through an attention mechanism. The invention utilizes multi-task learning of historical successful bridging patterns to generate atomic business instruction sequences optimally matched to the current business context. A dynamic risk verification module is configured to input the generated instruction sequences into a graph neural network-based discriminative model for multi-dimensional verification. By modeling the topological relationships between business objects, it predicts potential conflicts and compliance risks in the instruction execution path and outputs risk scores and optimization schemes. An adaptive execution and evolution module is configured to dynamically adjust the instruction execution strategy based on the risk score, execute verified instruction sequences, and construct triplet feedback data containing the instruction sequences, execution environment snapshots, and performance indicators. It continuously optimizes the parameter weights of the generation and discriminative models through an online learning mechanism. This invention provides an electronic device, including a processor, a memory, and a computer program stored in the memory, characterized in that the processor implements the above method when executing the program.

[0095] The present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that the program implements the above-described method when executed by a processor.

[0096] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A cross-platform business process automation bridging method for heterogeneous enterprise systems, characterized in that, The method comprises the following key steps: a multi-source context awareness step: real-time monitoring of a predefined business node event in a first business platform, in response to the event trigger, synchronously acquiring current state information, metadata and change history of a corresponding business object in a second business platform through a standardized interface, and collecting full-amount state context of the corresponding business object in the first business platform; An intelligent instruction generation step: multi-dimensional state features from the two platforms are constructed into a time-sequenced feature tensor, which is input into a sequence-to-sequence generation model based on a Transformer architecture; the model analyzes the semantic association between cross-platform states through an attention mechanism, generates an atomic business instruction sequence that is optimally matched with the current business context based on multi-task learning of historical successful bridging patterns; A dynamic A risk verification step: the generated instruction sequence is input into a discriminative model based on a graph neural network for multi-dimensional verification, the model models the topological relationship between business objects, predicts potential conflicts and compliance risks in the instruction execution path, and outputs a risk score and an optimization scheme; An adaptive execution and evolution step: dynamically adjusting the instruction execution strategy according to the risk score, executing the verified instruction sequence, and constructing a triple feedback data including the instruction sequence, the execution environment snapshot and the performance index, and continuously optimizing the parameter weights of the generation model and the discriminative model through an online learning mechanism.

2. The method of claim 1, wherein, The intelligent instruction generation step comprises the following sub-steps: a feature tensor construction sub-step: the state information from the two platforms is normalized, the classification features are mapped to dense vectors through an embedding layer, the numerical features are adaptively normalized by bucketing, and the time-sequenced features are extracted by a long short-term memory network to extract periodic pattern representation, and finally spliced to form a time-sequenced feature tensor; A hierarchical attention calculation sub-step: the generation model first analyzes the time sequence dependence of state changes within a single business platform through a platform-in attention layer using a causal mask mechanism, then calculates the association weight between the state features of the first business platform and the state features of the second business platform through a cross-platform attention layer using a cross-attention mechanism, and finally dynamically fuses the output results of the platform-in attention and the cross-platform attention through a gating mechanism; a multi-task instruction generation sub-step: the generation model simultaneously executes the main task to generate an atomic business instruction sequence, and predicts the state transition of the business platform at the next time through an auxiliary task.

3. The method of claim 1, wherein, The dynamic risk verification step comprises the following sub-steps: a business graph construction sub-step: taking a business object as a node and a business association between objects as an edge, a dynamic business graph with time sequence weight is constructed; a risk propagation analysis sub-step: the graph neural network adopts a time sequence graph attention network architecture, performs multi-hop information aggregation along the business graph through a message passing mechanism, preferentially analyzes the business objects on the key path dynamically identified by the PageRank algorithm, and calculates the cascading influence that may be caused by instruction execution; a multi-dimensional risk assessment sub-step: for the detected high-risk instruction sequence, an alternative path suggestion is automatically generated as an optimization scheme, and the business compliance, resource conflict probability and performance bottleneck risk scores of the instruction sequence are comprehensively evaluated to generate an overall risk score.

4. The method of claim 1, wherein, The adaptive execution and evolution step includes the following sub-steps: a policy selection sub-step: dividing the instruction sequence into three levels of high, medium and low risk according to the risk score, corresponding to three strategies of manual review and execution, restrictive automatic execution and full automatic execution respectively; a feedback data construction sub-step: recording the system state snapshots before and after the execution of the instruction sequence, collecting the execution time, resource consumption, business indicator change performance data, and constructing the triple feedback sample; An incremental learning optimization sub-step: selecting recent feedback data using a sliding window mechanism, setting a model update trigger to start model optimization when the time window set for balancing the real-time performance and stability of the model, the sample size threshold set to ensure sufficient training data, or the model performance degradation exceeds the pre-set threshold, using an elastic weight consolidation algorithm to prevent catastrophic forgetting, and coordinating the model updates of multiple business units through a federated learning architecture.

5. The method of claim 1, wherein, It also includes an elastic degradation step: continuously monitoring the service state and output confidence of the generated model; when detecting model service anomalies or confidence continuously below the threshold, automatically switching to a backup instruction generation mode based on a rule engine; the environmental data collected during the degradation period is quickly integrated into the training process through a priority playback mechanism after the model is restored.

6. A cross-platform business automation bridging system implementing the method of any of claims 1-5, wherein, It includes: A multi-source context awareness module configured to monitor predefined business node events in the first business platform in real time, and in response to event triggers, synchronously acquire the current state information, metadata and change history of the corresponding business objects in the second business platform through standardized interfaces, while collecting the full state context of the corresponding business objects in the first business platform; An intelligent instruction generation module configured to construct multi-dimensional state features from both platforms into a time-series feature tensor, input it into a sequence-to-sequence generation model based on the Transformer architecture, analyze the semantic association between cross-platform states through the attention mechanism, generate an atomic business instruction sequence that optimally matches the current business context based on multi-task learning of historical successful bridging patterns; a dynamic risk verification module configured to input the generated instruction sequence into a discriminative model based on graph neural networks for multi-dimensional verification, model the topological relationships between business objects to predict potential conflicts and compliance risks in the instruction execution path, and output a risk score and optimization scheme; An adaptive execution and evolution module configured to dynamically adjust the instruction execution strategy according to the risk score, execute the instruction sequence that passes the verification, and construct triple feedback data containing the instruction sequence, execution environment snapshot, and performance indicators, and continuously optimize the parameter weights of the generation model and discriminative model through online learning mechanism.

7. An electronic device comprising a processor, a memory, and a computer program stored on the memory, wherein the computer program, when executed by the processor, is arranged to perform the method of any one of claims 1 to 6. The processor executes the program to realize the method of any one of claims 1-5.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to realize the method of any one of claims 1-5. The program is executed by the processor to realize the method of any one of claims 1-5.

Images