Business modeling method based on AI model and service orchestration and computer program product

Abstract

The invention discloses a business modeling method based on an AI model and service orchestration and a computer program product, and the method comprises the steps: accessing multi-source heterogeneous data, and carrying out the standardization processing of the multi-source heterogeneous data; starting a multi-model collaboration process driven by a visual service orchestration engine, and distributing the standardized data to three AI models in parallel for screening, association, reasoning and prediction; meanwhile, configuring an aggregation node to perform weighted average algorithm fusion on output results of the three AI models to generate a prediction result; configuring a Bayesian regression node to carry out Monte Carlo sampling on a prediction result, and generating a demand distribution curve; and the configuration decision node generates an order based on a preset rule. Through the technical combination of visual arrangement, service grid and dynamic adaptation, seamless fusion of the AI capability and the service system is realized, a developer can intuitively combine the AI module and freely define the service logic flow without paying attention to a technical bottom layer, the development threshold is reduced, and the development efficiency and the system adaptability are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing systems or methods, and more specifically, relates to a business modeling method. Background Art

[0002] In recent years, the rapid development of artificial intelligence (AI) technology has promoted the popularization of the "Model as a Service (MaaS)" model. Developers can conveniently call pre-trained models (such as GPT, Stable Diffusion, etc.) through cloud platforms, significantly reducing the threshold for model development.

[0003] However, deeply integrating AI capabilities into specific business processes still faces many challenges, mainly reflected in: (1) Business logic often involves the coordinated invocation of multiple models (for example, after text generation, it is necessary to access the review model), the integration of third-party services (such as databases, payment interfaces), and the connection of manual review nodes. These links usually rely on manually writing code to implement service orchestration, which is a cumbersome process and has poor fault tolerance.

[0004] (2) Traditional service orchestration tools (such as Airflow, Kubeflow Pipelines) adopt a static configuration mode. When business rules change or models are iterated, the entire process needs to be redeployed, lacking the ability to adjust dynamically in real time.

[0005] (3) Developers need to master cross-domain skills such as model tuning, distributed communication protocols (such as gRPC, RESTful), and containerized deployment (such as Docker, Kubernetes) at the same time, thus raising the technical threshold for model development. Although low-code platforms (such as OutSystems, Mendix) attempt to simplify development through visual design, their functions mainly focus on general business processes. They lack both atomic encapsulation and dynamic scheduling support for AI models, and it is difficult to break through the limitations of fixed templates (such as complex conditional branches, loop controls, etc.). Moreover, some AI-specific tools (such as Google Vertex AI) are bound to specific cloud services or frameworks, resulting in insufficient scalability. Ultimately, code supplementation is still required to implement complex logic.

[0006] In response to the above technical problems, various solutions have been proposed at the present stage, but each has its own limitations. For example: The service orchestration engines of cloud providers (such as AWS Step Functions combined with SageMaker) call AI model services through JSON / YAML configuration. Although they support basic control flows, complex logic still needs to be supplemented by writing Lambda functions, and when business rules change, the entire workflow needs to be redeployed, unable to achieve dynamic hot updates.

[0007] Although low-code AI platforms (such as H2O.ai Driverless AI) provide visual model inference process design, they are limited to data science scenarios, difficult to deeply interact with external business systems (such as CRM, ERP), and have poor model compatibility.

[0008] Open-source workflow engines (such as Kubeflow Pipelines) build task dependencies based on DAGs, require developers to write Python code to handle node communication and parameter passing, and also need to operate and maintain the underlying cluster by themselves, with high development and operation and maintenance costs, and lack of visual collaboration capabilities.

[0009] The shortcomings of these solutions in terms of flexibility, ease of use, and real-time performance restrict the deep integration of AI and business systems. Summary of the Invention

[0010] The object of the present invention is to provide a business modeling method based on AI models and service orchestration to solve the problems of fragmented functions, technical coupling, and complex operation and maintenance of traditional solutions.

[0011] To solve the above technical problems, the present invention adopts the following technical solutions for implementation: In one aspect, the present invention proposes a business modeling method based on AI models and service orchestration, including the following processes: Access multi-source heterogeneous data and perform standardized processing on it; Start a multi-model collaboration process driven by a visual service orchestration engine and execute the following processes: Parallelly distribute the standardized data to three AI models, which are respectively used to screen, associate, infer, and predict the standardized data; Configure an aggregation node to fuse the output results of the three AI models using a weighted average algorithm to generate a prediction result; Configure a Bayesian regression node to perform Monte Carlo sampling on the prediction result to generate a demand distribution curve; Configure a decision-making node to generate an order based on preset rules.

[0012] In some embodiments of the present application, the process of accessing multi-source heterogeneous data and performing standardized processing on it may specifically include: Configure a data collection node and connect it to the enterprise system to extract the enterprise's historical data; Configure a logistics node to synchronously obtain external data streams; Configure a preprocessing node to automatically call the built-in Pandas engine to perform data cleaning.

[0013] In some embodiments of the present application, the data acquisition node can be configured to directly connect to an enterprise resource planning system (ERP) through the JDBC protocol to extract historical data of the enterprise; the historical data is structured feature data.

[0014] In some embodiments of the present application, the data cleaning process can be configured to include the following steps: Automatically identify missing values and fill the missing values using the moving window mean. Compress numerical features to the interval [0, 1] through the Min-Max normalization algorithm of the Scikit-learn library. Output a standardized JSON data packet carrying metadata tags.

[0015] In some embodiments of the present application, during the execution of the data cleaning, the sidecar proxy of the service mesh can be configured to intercept the original data stream in real time, and convert the CSV format into a Schema structure compatible with the downstream model node through a protocol bridge to ensure the consistency of the field order during model training.

[0016] In some embodiments of the present application, the three AI models can be encapsulated into three capability nodes and support cross-platform dynamic loading and invocation; among them, the three capability nodes can be respectively configured as: An LSTM neural network node, which focuses on capturing the non-linear associations between data, cascades SVR to optimize the residuals of the model, and analyzes the temporal dependence of the data. An ElasticNet regression node, which performs L1 regularization screening on sparse feature data to dynamically eliminate interference factors with a correlation lower than a set threshold. An XGBoost regression node, which performs inference and prediction based on the filtered feature dataset and outputs a trend prediction value.

[0017] In some embodiments of the present application, the set threshold can be dynamically adjusted according to the business scenario of the actual application on the visualization interface provided by the service orchestration to enhance the feature sparsity.

[0018] In some embodiments of the present application, in the multi-model collaboration process, the service mesh can be configured to continuously monitor the communication delay between each node. When the ElasticNet node times out due to data dimension mismatch, a circuit breaker mechanism is started, and the request is redirected to a pre-set random forest regression backup node to ensure that the business continuity is not affected by a single point of failure.

[0019] In another aspect, the present invention also proposes a computer program product, including a computer program or instruction, and when the computer program or instruction is executed by a processor, the following steps are performed: Access multi-source heterogeneous data and perform standardization processing on it; Start the multi-model collaboration process driven by the visualization service orchestration engine and execute the following procedures: Parallelly distribute the standardized data to three AI models, which are respectively used to screen, correlate, infer, and predict the standardized data; Configure an aggregation node to fuse the output results of the three AI models using a weighted average algorithm to generate a prediction result; Configure a Bayesian regression node to perform Monte Carlo sampling on the prediction result to generate a demand distribution curve; Configure a decision node to generate an order based on preset rules.

[0020] Compared with the prior art, the advantages and positive effects of the present invention are mainly reflected in: 1. Through the atomic encapsulation technology, the present invention uniformly abstracts AI models (such as XGBoost regression models, neural network models, etc.), business services (such as database reading and writing, APIs), and logical controls (such as branches, loops) into standardized nodes, encapsulating input and output interfaces, protocol conversion logics, and execution environment dependencies. Through metadata injection, "plug and play" across frameworks and protocols is achieved, solving the problem of manual adaptation between AI models and business services in traditional solutions.

[0021] 2. The present invention generates a flowchart that can be modified in real time based on a browser-side visualization editor (graphical interface), and the engine automatically converts the flowchart into a dynamic DAG (directed acyclic graph) execution plan.

[0022] 3. The present invention introduces a topology-aware scheduling algorithm, allowing dynamic adjustment of node attributes (such as switching model versions) and local reloading of subgraphs during runtime without downtime, solving the problem that traditional solutions (such as Airflow) must redeploy the entire process and lack the ability of real-time dynamic adjustment.

[0023] 4. Through the built-in service mesh, the present invention embeds a lightweight proxy (SidecarProxy) between nodes, which can automatically handle issues such as protocol conversion (gRPC↔REST), load balancing, and fault fusing between nodes, thus completely shielding the details of the underlying infrastructure.

[0024] 5. The present invention is compatible with heterogeneous models (such as TensorFlow, PyTorch, etc. models) and third-party APIs through a unified interface specification, supports private deployment and hybrid cloud architectures, and constructs an open ecosystem.

[0025] 6. Through the technical combination of "visual orchestration + service mesh + dynamic adaptation", the present invention effectively solves the problems of functional fragmentation, technical coupling, and complex operation and maintenance in traditional solutions, realizes the seamless integration of AI capabilities and business systems, significantly reduces the development threshold and improves the iteration efficiency, and provides a more agile and adaptable technical path for intelligent application development.

[0026] After reading the detailed description of the embodiments of the present invention in conjunction with the accompanying drawings, other features and advantages of the present invention will become clearer. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a general flowchart of an embodiment of the business modeling method based on AI models and service orchestration proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0028] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with 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 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.

[0029] In the description of the present invention, the consecutive numbering of method steps is for the convenience of review and understanding. Considering the overall technical solution of the present invention and the logical relationship between each step, adjusting the execution order between steps will not affect the technical effects achieved by the technical solution of the present invention.

[0030] In view of the problems faced by traditional AI application development technologies, such as high technical threshold, complex process and poor fault tolerance caused by the integration of AI models and business processes relying on manual coding, and lagging iteration response due to the lack of flexible adjustment ability of the system when business requirements change, this embodiment proposes a business modeling method based on AI models and service orchestration. It innovatively adopts a visual development mode, and through the deep integration of the graphical interface and service orchestration tools, developers can intuitively combine AI capability modules and freely define business logic flows without paying attention to the technical underlying layer, so as to quickly build scenario-based AI applications, significantly improving the development efficiency and system adaptability, and effectively reducing the application threshold of artificial intelligence technology.

[0031] Next, in combination with Figure 1 , the business modeling method of this embodiment will be elaborated in detail.

[0032] S110. Access multi-source heterogeneous data and perform standardization processing on it.

[0033] In this embodiment, this step may specifically include the following processes: S111. Configure data collection nodes and connect them to the enterprise system to extract historical data of the enterprise.

[0034] In this embodiment, the data collection nodes can be configured to directly connect to the Enterprise Resource Planning system (ERP) via the JDBC protocol to extract historical data of the enterprise.

[0035] The ERP system is a system for effectively sharing and utilizing enterprise resources. By fully organizing and effectively transmitting information, the enterprise's resources can be reasonably configured and utilized in various aspects such as procurement, inventory, production, sales, human resources, finance, and materials, thereby improving the enterprise's operating efficiency.

[0036] The historical data in this embodiment is structured feature data, and different historical data is extracted for different business scenarios. For example, in the application of supply chain demand forecasting, the historical sales data of the enterprise in the past three years can be extracted, including but not limited to structured fields such as SKU codes, quarterly sales volumes, and promotion activity markers.

[0037] S112. Configure logistics nodes to synchronously obtain external data streams.

[0038] In the application of supply chain demand forecasting, the external data streams can include supplier delivery cycles, raw material inventory levels, etc.

[0039] S113. Configure preprocessing nodes to automatically call the built-in Pandas engine to perform data cleaning.

[0040] In this embodiment, the data cleaning process can include: Automatically identify missing values and fill the missing values with the moving window mean; Compress numerical features to the interval [0, 1] through the Min - Max normalization algorithm of the Scikit - learn library; Output a standardized JSON data packet carrying metadata tags (such as data_type: structured, scaler: minmax).

[0041] During the data cleaning process, the sidecar proxy of the service mesh can be configured to intercept the original data stream in real - time, convert the CSV format to a Schema structure compatible with the downstream model nodes through a protocol bridge, and inject a feature name mapping table for the XGBoost regression node, for example, to ensure the consistency of the field order with that during model training.

[0042] S120. Start the multi - model collaboration process driven by the visualization service orchestration engine, screen, infer, and predict the standardized structured feature data, and generate orders.

[0043] In this embodiment, this step may specifically include the following processes: S121. Parallelly distribute the structured feature data after data cleaning to three AI models, encapsulate the three AI models into three capability nodes, and support cross-platform dynamic loading and invocation.

[0044] In this embodiment, the following three capability nodes can be configured: LSTM neural network node, which focuses on capturing the non-linear associations between data (for example, in supply chain demand forecasting applications, it can capture the non-linear associations between promotional activities and seasonal indices), cascades SVR to optimize the residuals of the model; at the same time, relying on the GPU acceleration container, analyze the time series dependence of data (such as sales data); and, built-in protocol converter, which can automatically convert JSON arrays into DMatrix objects for XGBoost model inference.

[0045] ElasticNet regression node, which performs L1 regularization screening on sparse feature data (such as raw material price fluctuation data) to dynamically eliminate interference factors with a correlation lower than the set threshold. The set threshold can be dynamically adjusted according to the business scenario of the actual application on the visual interface provided by service orchestration to enhance feature sparsity. In supply chain demand forecasting applications, the set threshold can be set to λ = 0.05.

[0046] XGBoost regression node, which performs inference and prediction based on the filtered feature set and outputs the trend prediction value; at the same time, configure the service mesh proxy of this node to immediately call the Tensor2JSON plugin to serialize it into a Base64-encoded string after detecting the PyTorch tensor output, and push it to the aggregation node after attaching dimension labels (such as "shape": [12, 64]).

[0047] S122. Configure the aggregation node to fuse the output results of the above three capability nodes using the weighted average algorithm to generate the prediction result.

[0048] S123. Configure the Bayesian regression node to perform Monte Carlo sampling on the prediction result to generate the demand distribution curve.

[0049] In this embodiment, the Bayesian regression node can be configured to generate a demand distribution curve with a 90% confidence interval.

[0050] S124. Configure the decision node to generate an order based on preset rules.

[0051] In supply chain demand forecasting applications, the decision node can be configured to call the procurement system API based on preset rules to generate an order.

[0052] In some embodiments, the preset rule can be configured as "trigger automatic procurement when the confidence interval width < 10%". If the uncertainty exceeds the limit, a to-do work order is pushed to the manual review module.

[0053] In the entire process S120 described above, the configured service mesh continuously monitors the communication delay between each node. When the ElasticNet node times out due to data dimension mismatch, the circuit breaker mechanism can be immediately activated, and the request is redirected to the pre-set random forest regression backup node to ensure that business continuity is not affected by a single point of failure.

[0054] In the retail price prediction scenario, the efficiency of the above technical solution of this embodiment is particularly remarkable, and the specific process is as follows: First, the ElasticNet regression node quickly filters out key features such as promotion activity intensity and competitor pricing changes from the real-time data stream. Its built-in L1 regularization coefficient is dynamically adjusted to λ = 0.1 through the visualization interface to enhance feature sparsity.

[0055] Then, the XGBoost regression node generates a benchmark price prediction value based on the filtered feature set, and through the cross-protocol routing function of the service mesh, transmits the prediction result to the Bayesian regression node in Protobuf format.

[0056] After that, the Bayesian regression node combines the historical price volatility data, quantifies the market risk through the Monte Carlo (MCMC) method, and finally outputs a dynamic price range (such as 199.99 ± 15.99).

[0057] After receiving the aggregation result, the decision node automatically calls the REST API of the ERP system to complete the price update, and preempts GPU resources for high-priority price adjustment tasks through the resource scheduling bus, reducing the end-to-end response time of the entire process from 6 hours of the traditional manual analysis method to 8 minutes.

[0058] In the above process, if the developer needs to urgently adjust the price strategy, the confidence threshold of the Bayesian node can be directly modified in the visual orchestration (for example, from 90% to 85%). The service mesh can complete the policy update within a short time (for example, within 5 seconds) by listening to the change events of the configuration center (Etcd). The newly initiated prediction requests take effect immediately, while the ongoing processes still use the old parameters until they end naturally, thus achieving seamless switching of business logic.

[0059] As can be seen from the technical implementation of the above scenario, the present application constructs an end-to-end intelligent pipeline covering data preprocessing, multi-model collaboration, and dynamic decision-making through the deep collaboration of atomic encapsulation, visual orchestration, and service mesh autonomy. At the resource scheduling level, the dynamic DAG engine of visual orchestration can automatically analyze and identify key path nodes (such as LSTM neural network nodes) based on the topological structure, and allocate exclusive GPU resources for them through the priority preemption mechanism of Kubernetes, reducing the neural network inference time by 58%. In terms of protocol compatibility, the converter function library built into the service mesh supports two-way conversion of data formats of more than 20 mainstream frameworks. The measured cross-framework communication latency is stable within 5ms, reducing the additional overhead by 92% compared with the manual coding scheme. These technologies transform AI application development from highly specialized coding work into business-oriented visual design, providing standardized and highly adaptable technical support for the digital transformation of industries such as retail, manufacturing, and healthcare.

[0060] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, for those of ordinary skill in the art, it is still possible to modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions required to be protected by the present invention.

Claims

1. A business modeling method based on an AI model and service orchestration, characterized in that, Including: Accessing multi-source heterogeneous data and performing standardization processing on it; Starting a multi-model collaboration process driven by a visualization service orchestration engine and performing the following processes: Parallelly distributing the standardized data to three AI models respectively for screening, associating, reasoning, and predicting the standardized data; Configuring an aggregation node to fuse the output results of the three AI models using a weighted average algorithm to generate a prediction result; Configuring a Bayesian regression node to perform Monte Carlo sampling on the prediction result to generate a demand distribution curve; Configuring a decision node to generate an order based on preset rules.

2. The business modeling method based on an AI model and service orchestration according to claim 1, wherein The process of accessing multi-source heterogeneous data and performing standardization processing on it includes: Configuring a data collection node and connecting it to the enterprise system to extract the enterprise's historical data; Configuring a logistics node to synchronously obtain external data streams; Configuring a preprocessing node to automatically call the built-in Pandas engine to perform data cleaning.

3. The business modeling method based on an AI model and service orchestration according to claim 2, wherein The data collection node is directly connected to the enterprise resource planning system ERP through the JDBC protocol, and the extracted enterprise historical data is structured feature data.

4. The business modeling method based on an AI model and service orchestration according to claim 2, wherein The data cleaning process includes: Automatically identifying missing values and filling the missing values using the moving window mean; Compressing numerical features to the range [0, 1] through the Min-Max normalization algorithm of the Scikit-learn library; Outputting a standardized JSON data packet carrying metadata tags.

5. The business modeling method based on an AI model and service orchestration according to claim 4, wherein During the process of performing the data cleaning, configure the sidecar proxy of the service mesh to intercept the original data stream in real time, and convert the CSV format to a Schema structure compatible with the downstream model nodes through a protocol bridge to ensure the consistency of the field order with that during model training.

6. The business modeling method based on AI models and service orchestration according to any one of claims 1 to 5, characterized in that Encapsulating the three AI models into three capability nodes for cross-platform dynamic loading and invocation; The three capability nodes are respectively: The LSTM neural network node, which focuses on capturing the non-linear associations between data, cascades SVR to optimize the residuals of the model, and analyzes the temporal dependence of the data; The ElasticNet regression node, which performs L1 regularization screening on sparse feature data to dynamically eliminate interference factors with a correlation lower than a set threshold; The XGBoost regression node, which performs reasoning and prediction based on the screened feature data set and outputs a trend prediction value.

7. The business modeling method based on an AI model and service orchestration according to claim 6, wherein According to the business scenario of actual application, dynamically adjust the set threshold on the visualization interface of the service orchestration to enhance feature sparsity.

8. The business modeling method based on an AI model and service orchestration according to claim 6, wherein In the multi-model collaboration process, configure the service mesh to continuously monitor the communication latency between each node. When the ElasticNet node times out due to data dimension mismatch, start a circuit breaker mechanism and redirect the request to a preset random forest regression backup node to ensure that the business continuity is not affected by a single point of failure.

9. A computer program product comprising a computer program / instructions, characterized in that, The computer program / instructions, when executed by a processor, implement the steps of the business modeling method based on AI models and service orchestration according to any one of claims 1 to 8.

Images