An ai intelligent decision optimization system for heterogeneous data

Abstract

The application belongs to the technical field of artificial intelligence, and particularly relates to an AI intelligent decision optimization system for heterogeneous data, which comprises a general heterogeneous data integration management module, a dynamic self-adaptive decision module, a closed-loop iterative optimization module and an intelligent computing power scheduling and fault tolerance module; the general heterogeneous data integration management module is used for innovatively aligning and fusing feature vectors of structured data, text and images, supporting both horizontal feature splicing based on entity ID and semantic relationship fusion based on a knowledge graph, which effectively breaks through data barriers, enables a decision model to simultaneously perform reasoning based on multi-dimensional fused features, significantly improves data utilization efficiency and lays a solid foundation for exploring implicit correlations of cross-modal data.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically referring to an AI intelligent decision optimization system for heterogeneous data. Background Technology

[0002] Various business scenarios involve heterogeneous data of multiple types, including structured data, text, and images. Traditional systems lack a unified data governance architecture, making it impossible to achieve standardized access and integration of multi-source data, easily forming data silos and wasting data resources. Simultaneously, traditional decision-making models have fixed algorithms and parameters, failing to adapt to the characteristics of heterogeneous data and dynamic business scenarios, making it difficult to uncover implicit data correlations, resulting in poor decision accuracy and targeting. Furthermore, the system lacks a closed-loop feedback iteration mechanism, leading to a continuous decline in decision performance over long-term operation. Additionally, its computational power scheduling is inefficient and its fault tolerance is weak, making it prone to response delays and decision failures in high-concurrency scenarios. Summary of the Invention

[0003] In order to overcome the shortcomings of the prior art, the present invention provides a digital twin remote operation and maintenance method and system for intelligent manufacturing.

[0004] The technical solution adopted by this invention is as follows: This invention provides an AI intelligent decision optimization system for heterogeneous data, comprising: A general heterogeneous data integrated governance module is used to access multi-source heterogeneous data, standardize the multi-source heterogeneous data, and generate a unified feature dataset through multi-modal feature fusion. A dynamic adaptive decision-making module, connected to the general heterogeneous data integrated governance module, includes: The scene recognition unit is used to identify the current business scene type based on the feature distribution of the feature dataset and business metadata; The algorithm matching unit is used to call the corresponding decision algorithm from a hybrid algorithm library containing at least two different types of algorithms based on the identified business scenario type. The parameter optimization unit is used to run the decision algorithm with a preset hyperparameter configuration when the decision is first executed, and to adjust the hyperparameters of the decision algorithm according to the decision performance feedback information after receiving the decision performance feedback information. The decision execution unit is used to execute decisions based on the adjusted hyperparameters and the feature dataset, and output the decision results. A closed-loop iterative optimization module, connected to the dynamic adaptive decision module, is used to collect feedback data on the decision result, calculate the decision deviation based on the feedback data and the decision result, provide the decision deviation as the decision performance feedback information to the parameter optimization unit, and update the model parameters of the decision algorithm in the dynamic adaptive decision module when the decision deviation meets the preset model update conditions. The intelligent computing power scheduling and fault tolerance module is connected to the general heterogeneous data integrated governance module, the dynamic adaptive decision-making module and the closed-loop iterative optimization module, respectively. It is used to monitor the system's computing power resources, dynamically allocate computing power resources according to the task priority and load status of each module, and perform service switching or degradation when a module instance failure is detected.

[0005] Furthermore, the universal heterogeneous data integrated governance module includes: A multi-source data access interface for receiving structured data, text data, and image data; Data standardization units are used to fill missing values ​​and standardize the format of structured data, segment and vectorize text data, and scale and normalize the pixels of image data to generate standardized structured records, text feature vectors and image feature vectors. The data fusion unit is used to horizontally concatenate the standardized structured records, text feature vectors, and image feature vectors based on entity association identifiers or timestamps to generate the unified feature dataset.

[0006] Furthermore, the scene recognition unit is equipped with a pre-trained scene classification model. The scene classification model takes the statistical features of the feature dataset and business metadata as input and outputs the scene type label of the current business scene. The statistical features include at least one of the feature mean, variance and missing rate.

[0007] Furthermore, the parameter optimization unit employs a reinforcement learning strategy to adjust hyperparameters; wherein, the reinforcement learning strategy uses the features of the feature dataset, the performance indicators of the decision algorithm, and the decision performance feedback information as states, the adjustment amount of hyperparameters as actions, and constructs a reward function with the goal of minimizing the decision bias, and trains a policy network to output online the adjustment actions of at least one hyperparameter among the learning rate, batch size, or regularization coefficient.

[0008] Furthermore, the closed-loop iterative optimization module includes: The feedback collection unit is used to collect user feedback on decision-making results or business performance evaluation data. The deviation calculation unit is used to convert the correction operation data or business effect evaluation data into a real label of the same dimension as the decision result, and calculate the cross-entropy loss or mean square error between the decision result and the real label as the decision deviation. The iterative update unit is used to trigger the update of the model parameters of the decision algorithm when the decision deviation exceeds a preset threshold for multiple consecutive periods or the accumulated amount of feedback data reaches a preset capacity.

[0009] Furthermore, when the iterative update unit triggers a model parameter update, it determines whether the current decision algorithm is a differentiable parameterized model. If so, it adopts an incremental training method, using only the newly added feedback data to perform mini-batch gradient descent updates on the original model parameters, and introduces a knowledge distillation loss term during the update process. The model before the update is used as the teacher model, and the model during the update is used as the student model, constraining the consistency of the distribution of their outputs. If the current decision algorithm is a non-parametric model, it updates by adjusting the rule weights or decision thresholds within the algorithm.

[0010] Furthermore, the hybrid algorithm library includes at least two of the following: deep learning models, tree models, and rule engines; the algorithm matching unit maintains a preset mapping table between scene types and algorithms. When the identified scene type is a high-complexity unstructured data scene, the deep learning model is matched first; when the identified scene type is a low-complexity structured data scene, the tree model or rule engine is matched first.

[0011] Furthermore, the intelligent computing power scheduling and fault tolerance module includes: The resource monitoring unit is used to collect the utilization of CPU, memory and GPU through agents deployed on computing nodes and store them in a time-series database; The computing power scheduling unit is used to maintain a task priority queue. When a high-priority decision task arrives and the current resource utilization exceeds the threshold, it automatically reduces the resource quota of low-priority batch data processing tasks and allocates the released resources to the high-priority decision task. The fault-tolerant processing unit is used to perform heartbeat detection on each module instance. When an instance failure is detected, the request traffic is switched to the backup instance, and a preset degradation strategy is enabled. The degradation strategy includes switching the time-consuming deep learning decision path to a lightweight tree model decision path.

[0012] Furthermore, the dynamic adaptive decision-making module and the closed-loop iterative optimization module communicate through an asynchronous message queue; the hyperparameter adjustment operation of the parameter optimization unit and the model parameter update operation of the iterative update unit are executed mutually exclusively. When the model parameter update is triggered, the parameter optimization unit pauses the hyperparameter adjustment. After the model parameter update is completed, the state of the parameter optimization unit is reset and initialized with the updated model parameters.

[0013] This solution also discloses an operation method for an AI-based intelligent decision optimization system for heterogeneous data, which mainly includes the following steps: Step A1: Access multi-source heterogeneous data through the general heterogeneous data integrated governance module, perform data standardization and multimodal feature fusion, and generate a unified feature dataset; Step A2: The dynamic adaptive decision module identifies the current business scenario type based on the feature dataset and business metadata, matches the corresponding decision algorithm from the hybrid algorithm library, and runs it with preset hyperparameters during the initial execution. After receiving decision performance feedback, it dynamically adjusts the hyperparameters and outputs the decision result based on the adjusted decision algorithm. Step A3: The closed-loop iterative optimization module collects user feedback data, calculates the decision deviation, provides the decision deviation as decision performance feedback to the dynamic adaptive decision module, and updates the model parameters of the decision algorithm when the model update conditions are met. Step A4: The intelligent computing power scheduling and fault tolerance module monitors computing power resources in real time, dynamically schedules resources according to task priority, and performs service switching and degradation processing when a fault is detected.

[0014] The beneficial effects achieved by the present invention using the above structure are as follows: The present invention provides an AI intelligent decision optimization system for heterogeneous data, achieving the following beneficial effects: (1) Through the general heterogeneous data integrated governance module, this application innovatively aligns and fuses the feature vectors of structured data, text and images, which supports both horizontal feature splicing based on entity ID and semantic relationship fusion based on knowledge graph. This effectively breaks down data barriers, enabling decision-making models to reason based on multi-dimensional fused features at the same time, significantly improving data utilization efficiency and laying a solid foundation for discovering implicit associations in cross-modal data.

[0015] (2) Through the dynamic adaptive decision-making module, this application establishes an intelligent decision-making link of scene recognition-algorithm matching-parameter optimization. The system can automatically identify business scenarios, call the most matching algorithm from the hybrid algorithm library as needed, and innovatively introduce a hyperparameter dynamic optimization mechanism based on reinforcement learning, so that the model parameters can be adjusted in real time as the data distribution changes. Compared with the traditional fixed single model, this solution greatly improves the pertinence and accuracy of decision-making.

[0016] (3) Through the closed-loop iterative optimization module, this application constructs a complete automated closed loop from decision output to feedback collection, then to deviation calculation and finally to model update. This enables the model to learn new knowledge fully while effectively suppressing the catastrophic forgetting of old knowledge, ensuring the continuous stability and improvement of the system's decision performance throughout its entire life cycle, and completely solving the problem of performance degradation after long-term model operation.

[0017] (4) Through intelligent computing power scheduling and fault tolerance modules, this application introduces a dynamic resource scheduling mechanism based on task priority, which automatically degrades non-core tasks in high-concurrency scenarios and prioritizes low-latency response for online decision-making. At the same time, automatic fault switching based on heartbeat detection and preset service degradation strategies ensure that the system can still maintain the availability of core decision-making functions when some components fail, and the overall architecture has extremely high robustness and business continuity assurance capabilities. Attached Figure Description

[0018] Figure 1 This is a flowchart of an AI intelligent decision optimization system for heterogeneous data proposed in this invention.

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] This invention provides an AI-powered intelligent decision-making optimization system for heterogeneous data. It employs a microservice architecture, with each functional module deployed as an independent service. Modules communicate with each other via HTTP / RPC protocol and asynchronous message queues. The system is divided into four core modules: a general heterogeneous data integrated governance module, a dynamic adaptive decision-making module, a closed-loop iterative optimization module, and an intelligent computing power scheduling and fault tolerance module.

[0022] I. General Heterogeneous Data Integrated Governance Module This module is responsible for solving the problem of unified access and fusion of multi-source heterogeneous data, providing a standardized data foundation for upper-level decision-making.

[0023] 1. Multi-source data access interface This interface supports access and adaptation to various data sources. For structured data, access is achieved by configuring JDBC connection information; for text data, streaming acquisition is supported via message pipelines such as Filebeat and Kafka; for image data, a REST API interface is provided for business systems to upload data, or batch import can be performed by connecting to a distributed file system.

[0024] 2. Data Standardization Unit After receiving multi-source data, the data standardization unit preprocesses different types of data separately: For structured data: Field mapping is performed based on a pre-defined data dictionary, unifying heterogeneous field names from different sources into standard field names. For missing values, numeric fields are filled with the median or mean, while categorical fields are filled with the mode or a special marker. Date formats are uniformly converted to "yyyy-MM-dd HH:mm:ss".

[0025] For text data: First, a general Chinese word segmentation tool is used for word segmentation and stop word removal. Then, a pre-trained BERT-base-chinese model is used, and the 768-dimensional output vector of the last layer [CLS] is taken as the semantic feature vector of the text.

[0026] For image data: the image is uniformly scaled to 224×224 pixels, and the pixel values ​​are normalized to the [0,1] range. Then, a ResNet-50 model pre-trained on the ImageNet dataset is used, and the 2048-dimensional vector output from its global average pooling layer is taken as the image feature vector.

[0027] 3. Data Fusion Unit The data fusion unit is responsible for associating and fusing standardized multimodal data, providing two optional solutions: Option 1: Horizontal Concatenation of Feature Vectors. For different modalities with clear entity associations or timestamp alignments, the corresponding structured records, text feature vectors, and image feature vectors are directly concatenated horizontally to form a unified fused feature vector. For example, the final feature vector dimension of a user entity is: structured feature dimension + 768 + 2048. This option is suitable for scenarios with clear data alignment relationships and has high computational efficiency.

[0028] Option 2: Knowledge Graph-Based Graph Structure Fusion. This option is used when the relationships between data are complex and require in-depth semantic mining. First, a domain knowledge graph is constructed, aligning entities from structured records and extracted text with nodes in the knowledge graph. Entity attributes are composed of feature vectors from their respective modalities. The fused data is stored in a graph structure, with entities as nodes and relationships as edges. Each node also carries the multimodal features of its associated nodes. This option preserves richer semantic relationships and is suitable for complex association analysis scenarios.

[0029] After generating a unified feature dataset, it is stored in a distributed feature storage system for use by downstream decision-making modules.

[0030] II. Dynamic Adaptive Decision-Making Module This module is the core decision engine of this invention, responsible for intelligently matching algorithms and dynamically optimizing parameters based on business scenarios to generate accurate decision results. To solve the cold start problem, the module has a clearly defined initialization and iterative operation mechanism.

[0031] 1. Scene Recognition Unit Internally, a pre-trained scene classification model based on a multilayer perceptron is deployed. The input vector of this model is composed of two parts: one is the statistical features of the current feature dataset; the other is the business metadata features. The model output is a probability distribution across preset business scene categories, and the category with the highest probability is taken as the label of the current business scene.

[0032] The scenario classification model is trained offline before being deployed online. The training data comes from the statistical features of historical business data and manually labeled scenario tags.

[0033] 2. Algorithm Matching Unit An internal scene-algorithm mapping table is maintained. This mapping table is pre-configured based on the data characteristics of each scene. For example: For scenarios with large amounts of data and complex features, such as "high-risk transaction detection", the data is mapped to a Transformer-based deep learning model, leveraging its powerful feature cross-referencing capabilities.

[0034] For scenarios of moderate complexity, such as "customer segmentation," which require a certain level of interpretability, the XGBoost tree model is used for mapping.

[0035] For scenarios with clear logic and extremely high real-time requirements, such as "simple rule verification", the rules are mapped to a rule engine based on Drools.

[0036] After receiving the scene labels output by the scene recognition unit, the algorithm matching unit determines the decision algorithm to be called for this decision task by looking up a table.

[0037] 3. Parameter Optimization Unit The parameter optimization unit employs a reinforcement learning algorithm based on proximal policy optimization to dynamically adjust the hyperparameters of the decision-making algorithm. This unit is crucial for addressing the issue of rigid decision models and adapting them to dynamic business requirements. Its core definition and workflow are as follows: Cold start mechanism: When the system starts for the first time or a new decision algorithm is matched without performance feedback data, the parameter optimization unit uses the default hyperparameter configuration preset by the algorithm to run the decision algorithm, ensuring that the system can be put into operation immediately.

[0038] Reinforcement learning component definition: Status: This includes the statistical distribution of the current feature dataset, the hyperparameter values ​​of the current decision algorithm, and the decision performance feedback values ​​received from the closed-loop iterative optimization module.

[0039] Actions: Adjustments to hyperparameters such as learning rate, batch size, and regularization coefficient; the output is a continuous adjustment amount.

[0040] Rewards: To minimize decision bias, the negative value of the decision bias is used as the reward. A resource consumption penalty is also introduced; if the adjusted hyperparameters lead to a significant increase in computation time, a negative reward is given, guiding the policy network to seek a balance between accuracy and efficiency.

[0041] Network structure: Both the policy network and the value network use three-layer fully connected neural networks.

[0042] Operational flow: After receiving the decision deviation from the closed-loop iterative optimization module, the parameter optimization unit uses it as the new state input to the policy network. The policy network outputs an action to update the hyperparameter configuration of the decision algorithm. The decision execution unit then uses the new configuration for inference.

[0043] 4. Decision-making and execution unit This unit loads the decision algorithm model adjusted by the parameter optimization unit, takes the unified feature dataset provided by the general heterogeneous data integration governance module 100 as input, performs model inference, and outputs the final decision result.

[0044] III. Closed-loop iterative optimization module This module is responsible for collecting feedback, quantifying biases, and driving continuous model updates to form a decision-making closed loop.

[0045] 1. Feedback Acquisition Unit Data points are embedded in the business front-end interface or back-end API logic to collect user feedback on decision results. For decision-making tasks, the corrected result value is recorded directly; for evaluation tasks, the user's actions are recorded. The collected data, along with the request ID and feature snapshot at the time of decision-making, is sent to a message queue (Kafka).

[0046] 2. Deviation Calculation Unit Feedback data is consumed from the message queue, and bias quantification is performed. For classification tasks, the user's correction result is used as the true label, and the decision result is used as the predicted value; the cross-entropy loss between the two is calculated. For regression tasks, the mean squared error is calculated. The calculated value is used as the decision bias for this round.

[0047] 3. Iterative update unit The decision bias is sent in real time to the parameter optimization unit in the dynamic adaptive decision module as a state and reward signal for its reinforcement learning. Simultaneously, feedback data and decision bias are stored in a circular buffer. When the preset model update trigger condition is met, the model parameters are updated. The update process is handled according to the algorithm type branch: For differentiable parameterized models such as deep learning, an incremental training approach is adopted. Specifically, a copy of the current model is first created as the teacher model, and its parameters are frozen. Then, the student model is updated using mini-batch gradient descent with new data from the buffer. When calculating the loss function, in addition to the standard task loss, a knowledge distillation loss term is introduced to calculate the KL divergence between the output probability distributions of the student and teacher models under the same input. By minimizing the total loss including the distillation loss, it is ensured that the model adapts to new data without forgetting old knowledge, thus avoiding catastrophic forgetting.

[0048] For non-parametric models such as tree models or rule engines: updates are achieved by adjusting rule weights or decision thresholds. For example, for a rule engine, the hit rate and false positive rate of each rule are statistically analyzed based on feedback data, and the priority of the rules or the threshold parameters are dynamically adjusted.

[0049] 4. Update the mutual exclusion mechanism To avoid conflicts between the online hyperparameter adjustments by the parameter optimization unit and the background model parameter updates by the iterative update unit, the system uses a distributed lock for mutual exclusion. When the iterative update unit initiates a model parameter update, it acquires a global update lock. Upon detecting this lock, the parameter optimization unit pauses hyperparameter adjustments. After the model parameter update is complete and the new model is loaded and deployed, the lock is released, the parameter optimization unit resets the policy network state, and restarts hyperparameter optimization based on the new model.

[0050] IV. Intelligent Computing Power Scheduling and Fault Tolerance Module This module is responsible for ensuring the stability of the system under high concurrency and abnormal conditions.

[0051] 1. Resource Monitoring Unit Deploy a resource collection agent on each service node to periodically collect metrics such as CPU, memory, GPU utilization, and network I / O, and store them in a time-series database.

[0052] 2. Computing power scheduling unit A global task priority queue is maintained, with high-priority tasks taking precedence over low-priority tasks. When the resource monitoring unit 410 detects that CPU or memory utilization exceeds the threshold of 80%, the computing power scheduling unit 420 reduces the number of Pod replicas for low-priority tasks via the Kubernetes API, prioritizing the release of computing resources to ensure the SLA of high-priority tasks. When resource utilization drops below 40%, low-priority tasks are gradually restored.

[0053] 3. Fault-tolerant processing unit Heartbeat checks are performed on each microservice instance; a failure is identified as a three-times unresponsive heartbeat. Upon failure, the failed instance is taken offline via the service registry, and the API gateway automatically switches traffic to healthy instances. Simultaneously, a preset degradation strategy is activated: if most instances of the dynamic adaptive decision-making module fail, the decision path is automatically downgraded from a complex deep learning model to a rule engine decision, and an alert is issued to operations personnel.

[0054] V. Overall System Workflow A complete decision-making process of the system of this invention is as follows: S1: Multi-source data are accessed through a general heterogeneous data integrated governance module, and standardized and integrated to generate a unified feature dataset.

[0055] S2: The scene recognition unit in the dynamic adaptive decision-making module identifies the current business scene; the algorithm matching unit matches the corresponding decision-making algorithm; the parameter optimization unit uses the default hyperparameters during the first run and inputs the feature data into the decision execution unit.

[0056] S3: The decision execution unit outputs the decision results to the business system.

[0057] S4: The feedback acquisition unit in the closed-loop iterative optimization module collects user feedback; the deviation calculation unit calculates the decision deviation; and the deviation is sent as performance feedback to the parameter optimization unit.

[0058] S5: The parameter optimization unit takes the bias as input and generates new hyperparameters through a reinforcement learning policy network for the next decision. Simultaneously, the iterative update unit incrementally trains the decision algorithm model using accumulated feedback data and updates it online when conditions are met.

[0059] S6: The entire process described above operates stably under the resource monitoring, elastic scheduling, and fault tolerance protection of the intelligent computing power scheduling and fault tolerance module.

[0060] VI. Hardware Implementation Environment This system can be implemented on general-purpose x86 architecture servers or cloud server clusters. Recommended single-node configuration: Intel Xeon 2.0GHz or higher CPU, 32GB or more of memory, and NVIDIA Tesla T4 processor for deep learning model training and inference. Software environment includes: Linux operating system (CentOS 7+), Python 3.8+, PyTorch 1.10+, and Kubernetes 1.20+. Those skilled in the art can deploy and run this system without creative effort by following this configuration guide and combining it with the technical solutions disclosed in this invention.

[0061] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An AI-powered intelligent decision-making optimization system for heterogeneous data, characterized in that: include: A general heterogeneous data integrated governance module is used to access multi-source heterogeneous data, standardize the multi-source heterogeneous data, and generate a unified feature dataset through multi-modal feature fusion. A dynamic adaptive decision-making module, connected to the general heterogeneous data integrated governance module, includes: The scene recognition unit is used to identify the current business scene type based on the feature distribution of the feature dataset and business metadata; The algorithm matching unit is used to call the corresponding decision algorithm from a hybrid algorithm library containing at least two different types of algorithms based on the identified business scenario type. The parameter optimization unit is used to run the decision algorithm with a preset hyperparameter configuration when the decision is first executed, and to adjust the hyperparameters of the decision algorithm according to the decision performance feedback information after receiving the decision performance feedback information. The decision execution unit is used to execute decisions based on the adjusted hyperparameters and the feature dataset, and output the decision results. A closed-loop iterative optimization module, connected to the dynamic adaptive decision module, is used to collect feedback data on the decision result, calculate the decision deviation based on the feedback data and the decision result, provide the decision deviation as the decision performance feedback information to the parameter optimization unit, and update the model parameters of the decision algorithm in the dynamic adaptive decision module when the decision deviation meets the preset model update conditions. The intelligent computing power scheduling and fault tolerance module is connected to the general heterogeneous data integrated governance module, the dynamic adaptive decision-making module and the closed-loop iterative optimization module, respectively. It is used to monitor the system's computing power resources, dynamically allocate computing power resources according to the task priority and load status of each module, and perform service switching or degradation when a module instance failure is detected.

2. The AI ​​intelligent decision optimization system for heterogeneous data according to claim 1, characterized in that: The universal heterogeneous data integrated governance module includes: A multi-source data access interface for receiving structured data, text data, and image data; Data standardization units are used to fill missing values ​​and standardize the format of structured data, segment and vectorize text data, and scale and normalize the pixels of image data to generate standardized structured records, text feature vectors and image feature vectors. The data fusion unit is used to horizontally concatenate the standardized structured records, text feature vectors, and image feature vectors based on entity association identifiers or timestamps to generate the unified feature dataset.

3. The AI ​​intelligent decision optimization system for heterogeneous data according to claim 2, characterized in that: The scene recognition unit is equipped with a pre-trained scene classification model. The scene classification model takes the statistical features of the feature dataset and business metadata as input and outputs the scene type label of the current business scene. The statistical features include at least one of the feature mean, variance and missing rate.

4. The AI ​​intelligent decision optimization system for heterogeneous data according to claim 3, characterized in that: The parameter optimization unit employs a reinforcement learning strategy to adjust hyperparameters. This strategy uses the features of the feature dataset, the performance metrics of the decision algorithm, and the decision performance feedback information as states, the adjustment amount of the hyperparameters as actions, and aims to minimize the decision bias to construct a reward function. It trains a policy network to output online adjustments to at least one hyperparameter among the learning rate, batch size, or regularization coefficient.

5. The AI ​​intelligent decision optimization system for heterogeneous data according to claim 4, characterized in that: The closed-loop iterative optimization module includes: The feedback collection unit is used to collect user feedback on decision-making results or business performance evaluation data. The deviation calculation unit is used to convert the correction operation data or business effect evaluation data into a real label of the same dimension as the decision result, and calculate the cross-entropy loss or mean square error between the decision result and the real label as the decision deviation. The iterative update unit is used to trigger the update of the model parameters of the decision algorithm when the decision deviation exceeds a preset threshold for multiple consecutive periods or the accumulated amount of feedback data reaches a preset capacity.

6. The AI ​​intelligent decision optimization system for heterogeneous data according to claim 5, characterized in that: When the iterative update unit triggers a model parameter update, it determines whether the current decision algorithm is a differentiable parameterized model. If so, it adopts an incremental training method, using only the newly added feedback data to perform mini-batch gradient descent updates on the original model parameters, and introduces a knowledge distillation loss term during the update process. The model before the update is used as the teacher model, and the model during the update is used as the student model, constraining the consistency of the distribution of their outputs. If the current decision algorithm is a non-parametric model, it updates by adjusting the rule weights or decision thresholds within the algorithm.

7. The AI ​​intelligent decision optimization system for heterogeneous data according to claim 6, characterized in that: The hybrid algorithm library includes at least two of the following: deep learning models, tree models, and rule engines. The algorithm matching unit maintains a preset mapping table between scene types and algorithms. When the identified scene type is a high-complexity unstructured data scene, the deep learning model is matched first. When the identified scene type is a low-complexity structured data scene, the tree model or rule engine is matched first.

8. The AI ​​intelligent decision optimization system for heterogeneous data according to claim 7, characterized in that: The intelligent computing power scheduling and fault tolerance module includes: The resource monitoring unit is used to collect the utilization of CPU, memory and GPU through agents deployed on computing nodes and store them in a time-series database; The computing power scheduling unit is used to maintain a task priority queue. When a high-priority decision task arrives and the current resource utilization exceeds the threshold, it automatically reduces the resource quota of low-priority batch data processing tasks and allocates the released resources to the high-priority decision task. The fault-tolerant processing unit is used to perform heartbeat detection on each module instance. When an instance failure is detected, the request traffic is switched to the backup instance, and a preset degradation strategy is enabled. The degradation strategy includes switching the time-consuming deep learning decision path to a lightweight tree model decision path.

9. The AI ​​intelligent decision optimization system for heterogeneous data according to claim 8, characterized in that: The dynamic adaptive decision-making module and the closed-loop iterative optimization module communicate through an asynchronous message queue; the hyperparameter adjustment operation of the parameter optimization unit and the model parameter update operation of the iterative update unit are mutually exclusive. When the model parameter update is triggered, the parameter optimization unit pauses the hyperparameter adjustment. After the model parameter update is completed, the state of the parameter optimization unit is reset and initialized with the updated model parameters.

10. An operation method for an AI intelligent decision optimization system for heterogeneous data, characterized in that; Operating the AI ​​intelligent decision optimization system for heterogeneous data as described in claim 9 mainly includes the following steps: Step A1: Access multi-source heterogeneous data through the general heterogeneous data integrated governance module, perform data standardization and multimodal feature fusion, and generate a unified feature dataset; Step A2: The dynamic adaptive decision module identifies the current business scenario type based on the feature dataset and business metadata, matches the corresponding decision algorithm from the hybrid algorithm library, and runs it with preset hyperparameters during the initial execution. After receiving decision performance feedback, it dynamically adjusts the hyperparameters and outputs the decision result based on the adjusted decision algorithm. Step A3: The closed-loop iterative optimization module collects user feedback data, calculates the decision deviation, provides the decision deviation as decision performance feedback to the dynamic adaptive decision module, and updates the model parameters of the decision algorithm when the model update conditions are met. Step A4: The intelligent computing power scheduling and fault tolerance module monitors computing power resources in real time, dynamically schedules resources according to task priority, and performs service switching and degradation processing when a fault is detected.

Images