An adaptive enterprise strategic decision management system and method

Abstract

This invention discloses an adaptive enterprise strategic decision-making management system and method. The system includes: a multi-source data acquisition and integration module, used to acquire multi-source business data in real time, and integrate and process the acquired data through a big data processing cluster based on a big data processing framework to output feature data; an intelligent decision-making model module, used to receive feature data, and use a cascaded fusion model including a random forest model and a neural network model for predictive analysis to generate market demand forecast results and intelligent resource allocation suggestions; a dynamic feedback control module, used to process data streams in real time using a stream processing engine, and based on a reinforcement learning algorithm, dynamically optimize and adjust the intelligent resource allocation suggestions according to the real-time business status to generate the final execution instruction; and a visualization and interaction module, used to display the decision-making path, key influencing factors, and real-time business indicators through visual charts, and provide multi-dimensional data analysis views.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an adaptive enterprise strategic decision-making management system and method. Background Technology

[0002] As enterprises increase their informatization and accelerate digital transformation, they generate massive amounts of business data in their daily operations. This data covers multiple dimensions, including user order data, inventory data, production data, market environment data, and social media interaction data. Traditional enterprise decision-making systems rely on experience-based decisions and fixed rule engines. However, as enterprises undergo digital transformation, they are gradually shifting to a digital technology and data-driven model. This model is increasingly revealing the following significant technical problems in a rapidly changing market: (1) Decision lag: Traditional digital systems cannot realize real-time dynamic processing of high-speed big data flows. Due to data processing delays, decisions are often based on outdated information, resulting in a serious lack of ability for enterprises to respond to market changes and miss business opportunities; (2) Lack of adaptability: Fixed rule engines are difficult to adapt flexibly to complex and ever-changing business environments, such as sudden market events or changes in consumer behavior. The system cannot optimize and adjust the rules based on historical experience, thus affecting the accuracy of decision-making. (3) Lack of transparency in the decision-making process: The decision-making logic is usually closed in a black box model, making it difficult for decision-makers to trace and understand the specific basis of the decision, key influencing factors and data flow, which reduces the credibility and auditability of the decision; (4) Low resource allocation efficiency: Due to the lack of a scientific data-driven resource allocation model, corporate resources (such as human, material and financial resources) are often unevenly allocated, resulting in low resource utilization, increased operating costs, and impact on overall efficiency.

[0003] To address the aforementioned technical problems, this invention proposes an adaptive enterprise strategic decision-making management system, aiming to overcome the shortcomings of existing technologies. Summary of the Invention

[0004] The purpose of this invention is to provide an adaptive enterprise strategic decision-making management system and method to overcome the shortcomings of existing technologies and solve problems such as slow response, lack of adaptability and opaque decision-making process in traditional decision-making systems, so as to realize intelligent, real-time and adaptive optimization of enterprise strategic decision-making.

[0005] To achieve the above objectives, the present invention provides an adaptive enterprise strategic decision-making management system, comprising: a multi-source data acquisition and integration module, used to collect multi-source business data, including internal and external enterprise data, in real time, and to integrate and process the collected data through a big data processing cluster based on a big data processing framework, outputting feature data; and an intelligent decision-making model module, connected to the multi-source data acquisition and integration module, used to receive the feature data and perform predictive analysis using a cascaded fusion model including a random forest model and a neural network model, generating market demand forecast results and intelligent resource allocation suggestions; wherein, the cascaded fusion model uses the random forest model as the output... The system uses the output as one of the input features of the neural network model and performs weighted fusion of the results of each sub-model to generate the intelligent resource configuration suggestion; the dynamic feedback control module is connected to the multi-source data acquisition and integration module and the intelligent decision model module, respectively, and is used to process the data stream in real time using the stream processing engine, and dynamically optimize and adjust the intelligent resource configuration suggestion according to the real-time business status based on the reinforcement learning algorithm to generate the final execution instruction; and the visualization interaction module is connected to the intelligent decision model module and the dynamic feedback control module, and is used to display the decision path, key influencing factors and real-time business indicators through visualization charts, and provide a multi-dimensional data analysis view.

[0006] The cascaded fusion model is specifically used for: First-level model processing: taking the feature data as input, performing prediction through the random forest model, and outputting a first prediction result and a feature importance vector; Second-level model processing: taking a first output feature set including the feature data, the first prediction result, and the feature importance vector as input, performing deep prediction through a neural network model, and outputting a second prediction result; Dynamic weighted fusion: dynamically calculating and allocating fusion weights according to the performance of the first-level model and the second-level model on the validation set, performing a weighted summation of the first prediction result and the second prediction result, and outputting the intelligent resource allocation suggestion.

[0007] The cascaded fusion model further includes an intermediate model processing stage located between the first-level model processing and the second-level model processing. This stage is used to take the first output feature set as input, perform prediction through a gradient boosting model, and output intermediate prediction results. The second-level model processing takes a second output feature set, including the first output feature set, the intermediate prediction results, and cross-combination features of the first prediction results and the intermediate prediction results, as input and performs deep prediction through a neural network model.

[0008] The dynamic feedback control module includes: a reinforcement learning unit, integrated into the stream processing engine, used to infer based on the current state after processing each micro-batch of data, select the optimal action, and generate the final execution instruction; and an exception handling unit, used to track abnormal orders or sudden events in real time through the complex event processing technology of the stream processing engine, and when a preset abnormal pattern is detected, to trigger a preset safety boundary rule to correct or overwrite the final execution instruction output by the reinforcement learning unit.

[0009] The reinforcement learning unit defines a state space, an action space, and a reward function. The state space is defined as a temporal feature vector of orders, inventory, production, and market demand extracted from the real-time data stream. The action space includes at least discretized or continuous instructions for production volume adjustment, inventory transfer, and price strategy change. The reward function is defined as a weighted combination based on profit, customer satisfaction, inventory turnover rate, and capacity utilization rate.

[0010] The adaptive enterprise strategic decision-making management system also includes a data synchronization and consistency assurance module connected to the dynamic feedback control module, which is used to establish a two-way data synchronization channel between the stream processing engine and the big data processing framework, and ensure data consistency through a two-phase commit protocol.

[0011] The multi-source data acquisition and integration module is also used to perform automated feature engineering in the big data processing framework. The automated feature engineering includes: extracting statistical features, time-series features, cross features and embedding features from the acquired data, and merging the extracted features.

[0012] The visualization interaction module includes: a decision interpretation unit, used to obtain decision path information and key influencing factors from the intelligent decision model module, and to display the decision path and the distribution of the key influencing factors in the decision path in a visualization manner; and a human-machine collaborative decision unit, used to display the intelligent resource allocation suggestions and their confidence levels, and to receive the interactive operations of the decision-maker, and to use the interactive operations as feedback data.

[0013] The adaptive enterprise strategic decision-making management system also includes a self-evolution mechanism module, used for: setting a set of key performance indicators and periodically evaluating the system's actual performance on each indicator; comparing the actual performance with preset target values, and triggering a model optimization process when the difference exceeds a threshold; in the optimization process, recording and comparing the performance of different model versions and algorithm architectures through a model management toolchain, automatically selecting the best-performing model for deployment, and realizing the dynamic updating and introduction of the intelligent decision-making model module; and when the model optimization process is triggered, exploring new feature types, new neural network model structures, or new model integration methods through automated machine learning technology, realizing the dynamic introduction of system functions and adaptive improvement of decision-making performance.

[0014] Another aspect of this invention provides an adaptive enterprise strategic decision-making management method, employing the aforementioned adaptive enterprise strategic decision-making management system, and comprising the following steps: a multi-source data acquisition and integration step, which collects multi-source business data, including internal and external enterprise data, in real time, and integrates and processes the collected data through a big data processing cluster based on a big data processing framework to output feature data; an intelligent decision-making step, which receives the feature data and performs predictive analysis using a cascaded fusion model comprising a random forest model and a neural network model to generate market demand forecast results and intelligent resource allocation suggestions; wherein the cascaded fusion model uses the output of the random forest model as one of the input features of the neural network model, and performs weighted fusion of the results of each sub-model to generate the intelligent resource allocation suggestions; a dynamic feedback control step, which uses a stream processing engine to process the data stream in real time, and based on a reinforcement learning algorithm, dynamically optimizes and adjusts the intelligent resource allocation suggestions according to the real-time business status to generate the final execution instruction; and a visualization and interaction step, which displays the decision path, key influencing factors, and real-time business indicators through visual charts and provides a multi-dimensional data analysis view.

[0015] As can be seen from the above technical solutions, the advantages of the present invention are: This invention constructs a complete closed-loop system for enterprise strategic decision-making through the coordinated operation of four modules: multi-source data acquisition, intelligent decision-making model, dynamic feedback control, and visual interaction. This solution achieves real-time integration of multi-source heterogeneous data, improves prediction accuracy through a cascaded fusion model, utilizes reinforcement learning for dynamic optimization and adjustment, and makes the decision-making process transparent and credible through visualization. It effectively solves the technical problems of decision lag, lack of adaptability, opaque processes, and low resource allocation efficiency in existing technologies.

[0016] This invention significantly improves the performance of the core decision-making module through multi-level technical optimization. The refined scheme of the cascaded fusion model employs a cascaded fusion of random forests and neural networks, introducing a gradient boosting model as an intermediate level and feature cross-combination, fully leveraging the advantages of multiple models to significantly improve prediction accuracy and model expressive power. The dynamic feedback control module achieves real-time dynamic adjustment through a reinforcement learning unit integrated into the stream processing engine, while an anomaly handling unit ensures system stability. Simultaneously, the specific definitions of the state space, action space, and reward function enable the reinforcement learning model to accurately adapt to enterprise decision-making scenarios, achieving multi-objective optimization. The data synchronization and consistency guarantee module ensures eventual data consistency through a bidirectional data synchronization channel and a two-phase commit protocol, providing accurate state data for dynamic feedback control. Automated feature engineering extracts and merges statistical, temporal, cross, and embedded features from the collected data, providing high-quality feature input for the intelligent decision-making model. The synergistic effect of these technical solutions enables the system to possess core decision-making capabilities of high-precision prediction, real-time dynamic adjustment, and high reliability.

[0017] This invention achieves breakthroughs in both interactive transparency and system evolution capabilities. The decision interpretation unit visualizes the decision-making path and key influencing factors, transforming the decision-making process from a "black box" to a "white box," thus enhancing decision credibility and auditability. The human-machine collaborative decision-making unit displays resource allocation suggestions and their confidence levels, and receives interactive operations from decision-makers as feedback data, combining machine intelligence with human experience to improve decision flexibility and user adoption rates. The self-evolution mechanism module triggers model optimization through periodic evaluation of key performance indicator sets, automatically compares different model versions and deploys the optimal model, while simultaneously exploring new features, architectures, and integration methods through automated machine learning, enabling the system to continuously learn and self-evolve, maintaining a long-term optimal state. These technical features collectively construct a transparent, reliable, and continuously optimizing intelligent decision-making system. Attached Figure Description

[0018] Figure 1 A module structure diagram of an adaptive enterprise strategic decision-making management system provided in an embodiment of the present invention; Figure 2 This is a core technical architecture diagram of an adaptive enterprise strategic decision-making management system provided in an embodiment of the present invention; Figure 3 This is a functional architecture diagram of an adaptive enterprise strategic decision-making management system provided in an embodiment of the present invention; Figure 4 A decision-making flowchart of an adaptive enterprise strategic decision-making management system provided in an embodiment of the present invention; Figure 5 for Figure 1 Another module structure diagram of the adaptive enterprise strategic decision-making management system provided in an embodiment of the present invention; Figure 6 A schematic diagram of the intelligent decision analysis platform provided by the adaptive enterprise strategic decision management system of the present invention; Figure 7 A schematic diagram of the transaction statistical analysis platform provided by the adaptive enterprise strategic decision-making management system of the present invention; Figure 8 A flowchart of an adaptive enterprise strategic decision-making management method provided in another embodiment of the present invention; In the attached figures, the following labels are used: 1- Adaptive Enterprise Strategic Decision-Making Management System; 10-Multi-source data acquisition and integration module; 11-Intelligent Decision Model Module; 12-Dynamic feedback control module; 120 - Reinforcement Learning Units; 121 - Exception Handling Unit; 122 - Cooperative control unit; 13-Visual Interactive Module; 130 - Decision Interpretation Unit; 131-Human-Machine Collaborative Decision-Making Unit; 14-Data synchronization and consistency assurance module; 15-Self-evolution mechanism module; 2- Adaptive corporate strategic decision-making management methods; Steps S20~S23. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0020] Figure 1 The module structure diagram of the adaptive enterprise strategic decision-making system (SDM) 1 provided in an embodiment of the present invention includes a multi-source data acquisition and integration module 10, an intelligent decision model module 11, a dynamic feedback control module 12, and a visualization interaction module 13.

[0021] Figure 2 This is the core technology architecture diagram of the Adaptive Enterprise Strategic Decision Management System 1. Figure 2 The module collaboration process (data flow and control flow) of the SDM system is demonstrated, and the core logic is as follows: Data entry point: Data from multiple sources is distributed to Flink stream processing (real-time path) and Spark batch processing (offline path) via Kafka message queue; Feature engineering: The results of real-time feature computation (Flink side) and batch feature engineering (Spark side) are stored in the feature storage; Decision reasoning: The unified reasoning service calls on the data stored in the feature store and combines it with the capabilities of the intelligent decision model module 11 to generate a decision; Execution and Feedback: After the decision is executed, a closed loop is formed through KPI monitoring → system evolution → model training → model repository → policy update. At the same time, the state synchronization mechanism ensures the consistency of state between modules. Dynamic control: CEP anomaly detection on the Flink stream processing side → real-time control → decision execution, linked with dynamic feedback control module 12 to achieve real-time optimization.

[0022] like Figure 3 As shown, this is a functional architecture diagram of an adaptive enterprise strategic decision-making management system 1 provided in an embodiment of the present invention. Figure 3 As shown, this adaptive enterprise strategic decision-making management system 1 is divided into four levels from a functional logic perspective: Data Acquisition and Integration Layer: Responsible for accessing multi-source business data such as user order data, inventory resource data, production capacity data, and market environment data, and performing preliminary integration processing; Intelligent processing and decision-making layer: Based on the Spark big data processing framework and Flink streaming engine, it performs in-depth processing and feature engineering on the collected data; AI Model and Optimization Layer: Deploy random forest models, neural network models, and reinforcement learning algorithms to achieve market demand forecasting, intelligent resource allocation, and dynamic optimization, and equip the model with a self-evolution engine to achieve continuous optimization. Application and Presentation Layer: Provides strategic decision visualization and real-time monitoring dashboards, supporting user interaction and multi-dimensional data analysis.

[0023] From a technical implementation perspective, the above four layers can be further refined into a six-layer architecture, and its specific technical components and interface specifications are shown in Table 1 below: Table 1 L0 Data source layer Multi-source data access API Gateway, Kafka Connect, IoT Hub REST / WebSocket / MQTT / API L1 Integrated batch processing layer Real-time / batch data processing Flink, Spark, Kafka Streams Avro / Protobuf Schema / SchemaRegistry L2 Feature Engineering and Storage Layer Feature extraction, storage, and management Feature Store, Spark ML, Hudi gRPC / REST API L3 Intelligent decision-making layer Model reasoning and optimization decision making TF Serving, MLflow, RayRLlib, DeepSeek TensorFlow Serving API / Flink ML / Scikit-learn L4 Dynamic Coordination Control Layer Real-time control and anomaly handling Flink CEP, Rule Engine Event-driven API / PyTorch L5 Service and Visualization Layer Service encapsulation, user interaction Spring Boot, Vue.js, ECharts GraphQL / REST API From a data processing logic perspective, the adaptive enterprise strategic decision-making management system 1 of this invention adopts a layered architecture design, mainly including a data source layer (L0), a stream-batch integrated processing layer (L1), a feature engineering and storage layer (L2), an intelligent decision-making layer (L3), a dynamic coordination and control layer (L4), and a service and visualization layer (L5)—a core framework of six layers. Each layer is related to... Figure 1 The correspondence between the functional modules shown is as follows: The data source layer (L0), the stream and batch processing layer (L1), and the feature engineering and storage layer (L2) together constitute the multi-source data acquisition and integration module 10, which is responsible for data access, processing and feature engineering; The intelligent decision layer (L3) corresponds to the intelligent decision model module 11, which is responsible for predictive analysis based on feature data and generating decision suggestions; The dynamic coordination control layer (L4) corresponds to the dynamic feedback control module 12, which is responsible for dynamically optimizing and adjusting the decision recommendations based on real-time data streams; The Service and Visualization Layer (L5) corresponds to the Visualization and Interaction Module 13, which is responsible for providing users with a display of the decision-making process and an interactive interface.

[0024] The adaptive enterprise strategic decision-making management system 1 of this invention adopts a unified data model design, and the data specification is implemented using Avro and Protobuf Schema combined with Schema Registry. The unified data model is defined as follows: # Unified Data Model Definition class DataSourceType(str, Enum): ERP = "erp" CRM = "crm" SCM = "scm" MARKET = "market" SOCIAL = "social" IOT = "iot" AI = "DS" class UnifiedDataModel(BaseModel): "Unified Data Model" metadata: Dict[str, str] = Field( default_factory=lambda: { "version": "2.0", "protocol": "avro", "timestamp": datetime.utcnow().isoformat() } ) source: DataSourceType entity_type: str # "order", "inventory", "production" entity_id: str attributes: Dict[str, any] features: Optional[List[float]] = None embeddings: Optional[List[float]] = None class Config: json_encoders = { datetime: lambda dt: dt.isoformat() } The adaptive enterprise strategic decision-making management system 1 of this invention adopts a "command-event-query" pattern to implement instruction control, and implements asynchronous instruction transmission and state synchronization functions based on event-driven architecture (EDA) and command-query responsibility separation (CQRS). A command base class is designed for command adjustment and processing. Relevant parameters are as follows: # Command-Event-Query Pattern Implementation T = TypeVar('T') @dataclass class Command: """Command Base Class""" command_id: str timestamp: datetime command_type: str payload: Dict[str, any] def to_kafka_message(self) -> bytes: return json.dumps({ "command_id": self.command_id, "type": self.command_type, "timestamp": self.timestamp.isoformat(), "payload": self.payload }).encode('utf-8') @dataclass class AdjustProductionCommand(Command): """Adjust XX command""" def __init__(self, product_id: str, adjustment: int, priority:int = 1): super().__init__( command_id=f"prod_adj_{datetime.utcnow().timestamp()}", timestamp=datetime.utcnow(), command_type="adjust_production", payload={ "product_id": product_id, "adjustment": adjustment, "priority": priority, "reason": "demand_forecast" } ) class CommandHandler(ABC, Generic[T]): """Abstract class for command handlers""" @abstractmethod async def handle(self, command: Command) -> T: pass...... class ProductionCommandHandler(CommandHandler[Dict]): """XX command handler""" async def handle(self, command: Command) -> Dict: if command.command_type == "adjust_production": return await self._adjust_production(command.payload) async def _adjust_production(self, payload: Dict) -> Dict: # Execute XX command to adjust logic return {"status": "success", "adjusted_amount": payload["adjustment"]} The multi-source data acquisition and integration module 10 is used to collect multi-source business data, including internal and external enterprise data, in real time. It then integrates and processes the collected data through a big data processing cluster based on a big data processing framework, outputting feature data. The multi-source data acquisition and integration module 10 is also used to perform automated feature engineering within the big data processing framework. Automated feature engineering includes extracting statistical features, time-series features, cross features, and embedded features from the collected data, and merging the extracted features.

[0025] Specifically, the multi-source data acquisition and integration module 10 collects multi-source business data in real time through various access methods such as API interfaces and message queues. Internal data sources include user order data and inventory data from Enterprise Resource Planning (ERP) systems, supply chain data from Supply Chain Management (SCM) systems, customer information data from Customer Relationship Management (CRM) systems, and manufacturing execution and production data from Manufacturing Execution System (MES). External data sources include market data obtained through API interfaces, social media data accessed through Kafka streams, and environmental data containing weather, policy, and business information. This data covers all key information dimensions required for enterprise strategic decision-making and is characterized by diverse formats and structures.

[0026] At the data processing level, the multi-source data acquisition and integration module 10 constructs a big data processing cluster based on Apache Spark 3.0+ to integrate and process the acquired multi-source business data. First, a unified data reading and storage management system is implemented through a stream-batch orchestrator, establishing a data warehouse to support subsequent data quality management and evaluation. Data quality management includes checking the timeliness, consistency, and importance of data, and possessing self-correcting capabilities, such as automatically repairing missing or abnormal data through preset rules to ensure the accuracy and reliability of input data.

[0027] In terms of feature engineering, the multi-source data acquisition and integration module 10 performs automated feature engineering in the Spark cluster, deeply processing the raw data. Specifically, this includes: extracting statistical features (such as mean, variance, quantiles, etc.) and time-series features (such as autocorrelation coefficient, trend strength, etc.) from time-series data; constructing cross-features by cross-combining features from different data sources (such as the interaction between product category and season); and generating embedded features (such as converting high-cardinality category features into low-dimensional dense vectors using Word2Vec or graph embedding techniques). After completing various feature extractions, all features are merged to form a wide-table feature set. To further optimize feature quality, the multi-source data acquisition and integration module 10 performs importance assessment and feature selection on the merged features, for example, using the random forest algorithm to calculate feature importance and filter feature subsets, ultimately outputting structured feature data for use by the intelligent decision-making model module 11.

[0028] The above processing can be expressed mathematically as follows: Let the data source set be... After feature engineering processing by the Spark big data processing cluster, the output feature vector is: , in, For feature vectors, This represents the i-th feature index.

[0029] The intelligent decision-making model module 11 is connected to the multi-source data acquisition and integration module 10. It is used to receive feature data and perform predictive analysis using a cascaded fusion model that includes a random forest model and a neural network model to generate market demand forecast results and intelligent resource allocation suggestions. The cascaded fusion model uses the output of the random forest model as one of the input features of the neural network model and performs weighted fusion of the results of each sub-model to generate intelligent resource allocation suggestions.

[0030] In its implementation, the intelligent decision-making model module 11 constructs intelligent decision-making models based on machine learning frameworks such as Scikit-learn (e.g., Scikit-learn 1.0+), TensorFlow (e.g., TensorFlow 2.5+), or PyTorch (e.g., PyTorch 1.8+). Traditional machine learning algorithms such as random forests are implemented using Scikit-learn, while deep learning algorithms are implemented using TensorFlow or PyTorch. To ensure the model's real-time performance and adaptability, the intelligent decision-making model module 11 constructs a dynamic real-time feature engineering pipeline and a data quality management pipeline, integrating AI real-time capture and processing functions. Specifically, it first uses the tsfresh tool to automatically extract time-series features and constructs a cross-feature creation and extraction pipeline to calculate feature importance (Top-K feature selection). It then uses the KS test to monitor changes in feature distribution for feature monitoring and drift detection, ensuring the stability and effectiveness of the model's input features.

[0031] In terms of model architecture, the intelligent decision-making model module 11 adopts a cascaded fusion strategy. First, a cascaded pipeline is constructed, sequentially passing through a random forest model → gradient boosting model → neural network model for prediction. Then, cascaded weighted fusion is performed to calculate the prediction confidence, dynamically adjust the fusion weights, and update the model. This process ensures that the model can adaptively optimize based on real-time performance.

[0032] Specifically, the cascaded structure adopts a path of random forest model → gradient boosting model → neural network model to achieve random forest feature extraction, deep neural network processing, AI module integration, LSTM temporal feature construction, dynamic attention fusion, and final output. Simultaneously, a stacked ensemble model is further constructed based on sklearn, TensorFlow, PyTorch, and Ray RLlib to dynamically calculate fusion weights and achieve dynamic adjustment of "better performance, higher weight" through a value propagation mechanism. Furthermore, this module integrates SHAP value calculation for model interpretability analysis, making the decision-making process more transparent and reliable.

[0033] The cascaded fusion model is specifically used for: First-level model processing: Taking feature data as input, a random forest model is used for prediction, outputting the first prediction result and a feature importance vector. Its mathematical expression is: , Where T is the number of decision trees. It is the t-th decision tree. This is the input feature matrix.

[0034] Intermediate-level model processing: Taking the first output feature set, including feature data, the first prediction result, and the feature importance vector, as input, a gradient boosting model is used for prediction, outputting intermediate prediction results. Its mathematical expression is: , in, , Let M represent the feature importance vector of the random forest, and let M represent the number of iterations to improve the ranking. It represents the weights of the m-th base learner.

[0035] The second-level model processing: Taking the second output feature set, which includes the first output feature set, intermediate prediction results, and cross-combination features of the first and intermediate prediction results, as input, a deep prediction is performed through a neural network model, outputting the second prediction result. Its mathematical expression is: , in, , and These are the weight matrix and bias vector of the Lth layer, respectively. It is an activation function.

[0036] Dynamic weighted fusion: Based on the performance of the first-level model, middleware model, and second-level model on the validation set, the fusion weights are dynamically calculated and assigned to each, and the prediction results of each sub-module are weighted and summed to output intelligent resource allocation suggestions. The fusion formula is:

[0037] in, The weights are dynamically adjusted based on model performance: for example, Other weights are calculated similarly.

[0038] in, , , These represent the root mean square error (RMSE) of the random forest model, gradient boosting model, and neural network model on the validation set, respectively. The better-performing model (i.e., the smaller the RMSE), the larger the fusion weight, thus making the final fusion result more accurate and reliable.

[0039] The core decision-making process of the intelligent decision-making model module 11 can also be represented as the following stages: Phase 1: ; Phase 2: ; Phase 3: ; Phase 4:

[0040] in: It is the mapping function of the random forest model. It is the feature importance vector of the random forest. It is the mapping function of the neural network model. Indicates the AI ​​participation vector. It is a reinforcement learning strategy network. This is the current state. It is the Softmax normalization function, and α, β, γ are dynamically adjusted fusion weights.

[0041] like Figure 4 As shown in the SDM decision flowchart, the decision sub-process of module 11 of the intelligent decision model is as follows: Multi-source data acquisition: Integrates multi-source data such as orders, inventory, and market data; Spark data processing + feature engineering: cleaning, transforming, and extracting features from multi-source data; Intelligent decision-making model processing: Invoking random forest and neural network models, fusing the results to generate decision suggestions; Decision output: The decision recommendations are transmitted to the dynamic feedback control module 12; Flink Real-Time Feedback: Feeds the effects of decision execution back to the model, enabling continuous evolution through "model optimization feedback".

[0042] The dynamic feedback control module 12 is connected to the multi-source data acquisition and integration module 10 and the intelligent decision model module 11, respectively. It is used to process data streams in real time using a stream processing engine, and based on a reinforcement learning algorithm, dynamically optimize and adjust intelligent resource configuration suggestions according to real-time business status to generate final execution instructions. The dynamic feedback control module 12 includes a reinforcement learning unit 120 and an exception handling unit 121.

[0043] The reinforcement learning unit 120 defines a state space, an action space, and a reward function. The state space is defined as a temporal feature vector of orders, inventory, production, and market demand extracted from the real-time data stream. The action space includes at least discretized or continuous instructions for production volume adjustment, inventory transfer, and price strategy change. The reward function is defined as a weighted combination based on profit, customer satisfaction, inventory turnover rate, and capacity utilization rate.

[0044] In its implementation, the dynamic feedback control module 12 builds a real-time data processing pipeline based on the Apache Flink stream processing engine to achieve a rapid system response mechanism. By designing an adaptive control algorithm based on reinforcement learning, it enables self-iterative optimization of resource allocation strategies and service priority levels, as well as functions such as abnormal order tracking.

[0045] The reinforcement learning unit 120 is integrated into the Flink stream processing engine. It is used to infer based on the current state after processing each micro-batch of data, select the optimal action, and generate the final execution instruction.

[0046] State Space Definition: The state space comprises multi-dimensional elements extracted from the real-time data stream. Specifically, a state definition is constructed, including 23 dimensions such as orders, inventory, production, market, finance, and time, followed by the construction of transformation vectors. The state vectors cover: order volume, order amount, order fulfillment rate, number of delayed orders; current inventory, in-transit inventory, safety stock, inventory turnover days; equipment utilization rate, capacity load rate, work-in-process quantity, production yield; search popularity, competitor prices, market share; gross profit margin, cash flow; day of the week, whether it is a holiday, quarter, etc.

[0047] Action space definition: Action design is created from state vectors. Continuous actions include production adjustments (range -1.0 to 1.0, corresponding to adjustments from -100% to +100%), inventory adjustments, pricing strategies, resource allocation, etc. Based on this, 11 categories of discretized action vectors are constructed: maintain production, significantly increase production, slightly increase production, significantly decrease production, slightly decrease production, stockpiling, strong promotion, moderate promotion, price adjustment, resource allocation, and supply chain selection optimization.

[0048] Reward function design: Define 23 target values, set penalty coefficients, and calculate reward values. The reward function is a weighted combination of multiple objectives, including profit rewards, customer satisfaction rewards, inventory efficiency rewards, capacity utilization rewards, and stability rewards.

[0049] Learning Algorithm: The Q-learning algorithm is used for policy optimization, and its update formula is as follows:

[0050] Where s is the current state, a is the action to be performed, and r is the immediate reward. It's the learning rate. It is a discount factor.

[0051] Online Learning and Real-Time Control: Flink reinforcement learning ensemble is used to create online learning and real-time control pipelines. After processing each micro-batch of data, the reinforcement learning unit infers based on the current state, selects the optimal action, and generates the final execution instruction. Simultaneously, empirical data is stored in a replay buffer for periodically updating the Q-network parameters.

[0052] The exception handling unit 121 is used to track abnormal orders or sudden events in real time using Flink's Complex Event Processing (CEP) technology. When a preset exception pattern is detected (such as order volume continuously exceeding a threshold, inventory falling below safety stock, equipment failure signal, etc.), a preset safety boundary rule is triggered to correct or overwrite the final execution instruction output by the reinforcement learning unit.

[0053] Specifically, an exception handling and security boundary control mechanism is constructed, real-time control is enhanced using a security mechanism manager, security rules, boundary restriction rules, and check conditions are loaded, security correction rules are applied, and a log rolling replay mechanism is implemented. For example, when it is detected that the inventory of a certain SKU is lower than the safety stock and replenishment is delayed, the security boundary rule forcibly limits the production adjustment of the product to no more than ±20% and triggers a replenishment instruction.

[0054] To enhance the system's ability to coordinate with the external environment, the dynamic feedback control module 12 may also include a collaborative control unit 122, which is used to build an upstream and downstream collaborative mechanism. An adaptive feedback loop is embedded through the system integration manager to dynamically adjust resource allocation strategies and service priority levels based on real-time business data streams.

[0055] The collaborative control unit 122 includes upstream and downstream interface adapters for establishing data connections with external systems such as supplier management systems, customer relationship management systems, and logistics management systems. It acquires real-time data on upstream and downstream inventory status, order progress, and logistics information through standard API interfaces. The collaborative control unit 122 also includes an adaptive feedback regulator that calculates collaborative optimization coefficients based on resource allocation suggestions output by the reinforcement learning unit and combined with real-time upstream and downstream status data. When upstream supply delays or downstream demand surges are detected, a dynamic adjustment mechanism is automatically triggered to recalculate resource allocation priorities and generate collaborative adjustment instructions.

[0056] For example, in a manufacturing scenario, when upstream raw material suppliers experience supply delays, the collaborative control unit automatically adjusts the production schedule, prioritizing the allocation of limited raw materials to high-priority orders, while simultaneously notifying downstream customers of changes in expected delivery times.

[0057] like Figure 5 As shown, the adaptive enterprise strategic decision management system 1 also includes a data synchronization and consistency assurance module 14 connected to the dynamic feedback control module 12, which is used to establish a two-way data synchronization channel between the stream processing engine and the big data processing framework, and ensure data consistency through a two-phase commit protocol.

[0058] Specifically, a bidirectional synchronous data processing mechanism is established between the Flink stream processing engine and the Spark big data processing framework, namely Flink→Spark synchronization and Spark→Flink synchronization. A two-phase commit protocol (such as an implementation based on IcebergSink) is used to ensure eventual data consistency and provide accurate state data for the dynamic feedback control module 12.

[0059] The visualization interaction module 13, connected to the intelligent decision-making model module 11 and the dynamic feedback control module 12, is used to display the decision-making path, key influencing factors, and real-time business indicators through visual charts, and to provide multi-dimensional data analysis views. The visualization interaction module 13 includes a decision interpretation unit 130 and a human-machine collaborative decision-making unit 131.

[0060] In its implementation, the visualization interaction module 13 utilizes UI / UX design to create an integrated user interface, integrates ECharts to visualize decision tree data, and provides an intuitive interface for data analysis and decision support.

[0061] Decision interpretation unit 130: This unit acquires decision path information and key influencing factors from the intelligent decision model module 11, and visualizes the decision path and the distribution of key influencing factors within it. Specifically, it constructs a decision tree data visualization function to display the decision path and key influencing factors; it constructs a decision process transparency module to create a decision interpreter, generate and publish decision summaries, track decision paths, identify key factors, analyze confidence levels to find sources of uncertainty, assess impacts, and propose alternative solutions.

[0062] Human-machine collaborative decision-making unit 131: This unit displays intelligent resource allocation suggestions and their confidence levels, and receives interactive operations from decision-makers, using these operations as feedback data. Specifically, it utilizes AI technology to achieve human-machine collaborative decision-making, selects trusted data sources to construct a collaborative model, sets risk and credit levels, and forms a joint decision-making model combining autonomous decision-making and manual review, providing an intelligent decision recommendation system based on trust scores.

[0063] In addition, the visualization and interaction module 13 also constructs a real-time business indicator monitoring panel to dynamically display sales, inventory, and production data; it constructs a multi-dimensional data analysis view, supporting drill-down analysis based on time, product, region, and other dimensions. Simultaneously, it constructs a decision control panel, decision tree panel, user feedback panel, real-time monitoring panel, decision impact panel, system status panel, and decision log panel. It calls callback functions to update decision visualizations, updates real-time data processing user feedback, creates KPI dashboards to implement system prediction functions, and creates feature importance charts to analyze decision features.

[0064] The adaptive enterprise strategic decision-making management system 1 of this invention employs multiple types of databases to construct feature engineering and storage layers, providing data persistence, caching acceleration, and knowledge graph query services for each functional module. Specifically: MySQL 8.0 is used to store structured business data, query decision support data, and create decision logs; Redis 6.0+ is used to cache real-time data, creating a Redis cache manager to cache real-time data, obtain real-time metrics, obtain and cache model prediction results, and publish decision events; Neo4j 4.0+ is used to store strategic knowledge graphs, constructing strategic knowledge graphs, creating strategy-goal-resource sequences, querying strategic paths, analyzing strategic impacts, and identifying conflicting strategies to achieve dynamic optimization decision-making within the system.

[0065] As an optional enhancement module of the Adaptive Enterprise Strategic Decision Management System 1, the Strategic Knowledge Graph Management Module is used to define and manage enterprise strategy and construct an enterprise strategic knowledge graph. This module includes: a strategic element extraction unit, used to extract key strategic elements from enterprise strategic documents; a knowledge graph construction unit, which constructs an enterprise strategic knowledge graph based on a graph database (such as Neo4j) and stores the relationships between strategic elements; and a strategic context management unit, which configures differentiated decision rules for different strategic contexts, which are then used as constraints by the intelligent decision-making model module when generating resource allocation recommendations. This strategic knowledge graph management module works in conjunction with the Neo4j database used for data storage, providing strategic knowledge support for the system.

[0066] like Figure 5 As shown, the adaptive enterprise strategic decision-making management system 1 also includes a self-evolution mechanism module 15, used to achieve continuous learning and adaptive optimization of the system, maintaining a long-term optimal state. This self-evolution mechanism module 15 is connected to the intelligent decision-making model module 11, periodically evaluating system performance and triggering model updates. Specifically, it includes: KPI Setting and Evaluation: Define a set of key performance indicators, including prediction accuracy, response time (ms), resource utilization, user satisfaction, decision quality, and system stability, and calculate a comprehensive score. Use dictionary principles to standardize model performance recording, build a system evolution engine using MLflow, and periodically evaluate system performance using KPIs. Calculate prediction accuracy, response time, resource utilization, user satisfaction, decision quality, and system stability, and calculate the gaps to determine whether to implement evolution.

[0067] Model optimization process: When the actual performance deviates from the preset target value by more than a threshold, the model optimization process (including emergency adaptation) is triggered. In the optimization process, the performance of different model versions and algorithm architectures is recorded and compared through the model management toolchain (MLflow), and the model with the best performance is automatically selected for deployment, realizing the dynamic update and introduction of the intelligent decision-making model module 11.

[0068] Automated Machine Learning Exploration: When triggering the model optimization process, automated machine learning techniques are used to explore new feature types, new neural network model structures, or new model ensemble methods. Specifically, this includes: optimizing feature engineering and further exploring new algorithms, optimizing model architecture, and improving ensemble methods (such as using weighted voting ensemble, Bayesian averaging, etc.). The improved model is then deployed, adjusted, optimized, and stabilized, ultimately achieving full deployment and beginning the collection of new training data, thereby enabling the system's self-evolution.

[0069] The difference between the self-evolution mechanism module 15 and the dynamic feedback control module 12 is that the self-evolution mechanism module 15 performs offline or periodic model updates (such as daily or weekly), and achieves long-term performance optimization by retraining the model, searching for new features or new architectures; while the dynamic feedback control module 12 performs real-time or near real-time policy adjustments (second-level / minute-level). The two work together to ensure the system's short-term response capability and long-term evolution capability.

[0070] This adaptive enterprise strategic decision-making management system 1 also employs a dual-loop adjustment mechanism, including: developing an individual adaptive loop to achieve parameter self-adjustment of individual decision nodes; and constructing an organizational dynamic restructuring loop strategy to adjust the organizational structure of decision nodes according to strategic changes. This mechanism works in conjunction with the self-evolution mechanism module 15 to enhance the system's adaptability at both the individual and organizational levels.

[0071] To facilitate understanding of the above embodiments, a specific application scenario of the above embodiments will be used as an example for illustration below. Taking the production resource optimization and allocation of a manufacturing enterprise as an example, the specific implementation process and technical effects of the adaptive enterprise strategic decision-making management system 1 of the present invention will be explained: (1) Data collection: The system collects multi-source business data in real time through API interface, including: more than 5,000 user order data per day, inventory data of more than 10,000 SKUs, real-time status data of more than 200 production equipment, and market data from more than 10 data sources.

[0072] (2) Decision processing: The product demand for the next week was predicted using a random forest model with a prediction accuracy of 88%; the production schedule was optimized using a neural network model, which increased the equipment utilization rate from 65% to 82%.

[0073] (3) Dynamic optimization: By using the Flink stream processing engine to detect abnormal orders in real time, the average response time is reduced from 2 hours in the traditional system to 15 minutes; by using reinforcement learning algorithms to dynamically adjust resource allocation strategies, the inventory turnover rate is increased by 25%.

[0074] (4) Visualization: ECharts generates multi-dimensional data dashboards to help managers grasp the business status in real time and support rapid decision-making. For example... Figure 6 and Figure 7 As shown, the visualization and interaction module 13 provides an intelligent decision analysis platform and a transaction statistics analysis platform. The intelligent decision analysis platform displays the following: system overview indicators, annual sales trend forecast, current inventory level, equipment utilization rate, model performance comparison, feature importance analysis, real-time data monitoring, abnormal order monitoring, resource allocation and usage, production decision tree, and product sales distribution. The transaction statistics analysis platform displays the following: real-time transaction statistics for today's equipment spare parts, nuclear power plant conventional island equipment spare parts, thermal power plant equipment spare parts, and steel plant equipment spare parts; transaction statistics for this week; transaction statistics for this month; historical transaction trends (monthly statistics); and historical transactions and future forecasts (for the next two quarters).

[0075] Compared with the prior art, the present invention has the following significant advantages: (1) Technological innovation: For the first time, random forest, neural network and reinforcement learning algorithms are integrated and applied to corporate strategic decision-making to achieve multi-level intelligent decision-making; (2) System integrity: Provides a full-chain solution from data acquisition, processing, decision-making to visualization, covering the entire enterprise decision-making process; (3) Real-time performance and adaptability: Real-time response and continuous self-optimization of the system are achieved through Flink stream processing and reinforcement learning; (4) Scalability: It adopts a modular design and can flexibly expand functional modules according to enterprise needs; (5) User-friendliness: Through an intuitive visual interface, the threshold for use is lowered and the transparency of decision-making is improved.

[0076] The following are method embodiments corresponding to the system embodiments described above. These embodiments can be implemented in conjunction with the above embodiments. The relevant technical details mentioned in the above embodiments remain valid in these embodiments, and will not be repeated here to avoid repetition. Correspondingly, the relevant technical details mentioned in these embodiments can also be applied to the above embodiments.

[0077] like Figure 8 As shown, another embodiment of the present invention provides an adaptive enterprise strategic decision management method 2, which adopts the above-described adaptive enterprise strategic decision management system 1 and includes the following steps: Multi-source data acquisition and integration step S20: Real-time acquisition of multi-source business data, including internal and external enterprise data, and integration and processing of the acquired data through a big data processing cluster based on a big data processing framework to output feature data; Intelligent decision-making step S21: Receive feature data and perform predictive analysis using a cascaded fusion model that includes a random forest model and a neural network model to generate market demand forecast results and intelligent resource allocation suggestions; wherein, the cascaded fusion model uses the output of the random forest model as one of the input features of the neural network model, and performs weighted fusion of the results of each sub-model to generate the intelligent resource allocation suggestions; Dynamic feedback control step S22: The data stream is processed in real time using a stream processing engine, and the intelligent resource configuration suggestions are dynamically optimized and adjusted based on the real-time business status using a reinforcement learning algorithm to generate the final execution instruction; Visual interaction step S24: Display the decision-making path, key influencing factors and real-time business indicators through visual charts, and provide multi-dimensional data analysis views.

[0078] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms fall within the scope of protection of the present invention.

Claims

1. An adaptive enterprise strategic decision-making management system, characterized in that, include: The multi-source data acquisition and integration module is used to collect multi-source business data, including internal and external data, in real time, and integrate and process the collected data through a big data processing cluster based on a big data processing framework to output feature data. The intelligent decision-making model module, connected to the multi-source data acquisition and integration module, is used to receive the feature data and perform predictive analysis using a cascaded fusion model that includes a random forest model and a neural network model to generate market demand forecast results and intelligent resource allocation suggestions. The cascaded fusion model uses the output of the random forest model as one of the input features of the neural network model and performs weighted fusion of the results of each sub-model to generate the intelligent resource allocation suggestions. The dynamic feedback control module is connected to the multi-source data acquisition and integration module and the intelligent decision model module, respectively. It is used to process the data stream in real time using the stream processing engine, and dynamically optimize and adjust the intelligent resource configuration suggestions based on the real-time business status according to the reinforcement learning algorithm, and generate the final execution instruction. The visualization interaction module, connected to the intelligent decision-making model module and the dynamic feedback control module, is used to display the decision-making path, key influencing factors and real-time business indicators through visual charts, and to provide multi-dimensional data analysis views.

2. The adaptive enterprise strategic decision-making management system according to claim 1, characterized in that, The cascaded fusion model is specifically used for: First-level model processing: Taking the feature data as input, the random forest model is used for prediction, and the first prediction result and feature importance vector are output. Second-level model processing: Taking the first output feature set, which includes the feature data, the first prediction result, and the feature importance vector, as input, a deep prediction is performed through a neural network model, and a second prediction result is output. Dynamic weighted fusion: Based on the performance of the first-level model and the second-level model on the validation set, the fusion weights are dynamically calculated and assigned respectively. The first prediction result and the second prediction result are weighted and summed to output the intelligent resource configuration suggestion.

3. The adaptive enterprise strategic decision-making management system according to claim 2, characterized in that, The cascaded fusion model also includes an intermediate model processing level located between the first-level model processing and the second-level model processing, which is used to take the first output feature set as input, make predictions through a gradient boosting model, and output intermediate prediction results. The second-level model processing takes the first output feature set, the intermediate prediction result, and the cross-combination features of the first prediction result and the intermediate prediction result as input, and performs deep prediction through a neural network model.

4. The adaptive enterprise strategic decision-making management system according to claim 1, characterized in that, The dynamic feedback control module includes: The reinforcement learning unit, integrated into the stream processing engine, is used to infer based on the current state after processing each micro-batch of data, select the optimal action, and generate the final execution instruction. The exception handling unit is used to track abnormal orders or sudden events in real time through the complex event processing technology of the stream processing engine. When a preset exception pattern is detected, a preset security boundary rule is triggered to correct or overwrite the final execution instruction output by the reinforcement learning unit.

5. The adaptive enterprise strategic decision-making management system according to claim 4, characterized in that, The reinforcement learning unit defines a state space, an action space, and a reward function, wherein... The state space is defined as a temporal feature vector of orders, inventory, production, and market demand extracted from the real-time data stream. The action space includes at least discrete or continuous instructions for production volume adjustment, inventory transfer, and price strategy change. The reward function is defined as a weighted combination of profit, customer satisfaction, inventory turnover rate, and capacity utilization rate.

6. The adaptive enterprise strategic decision-making management system according to claim 1, characterized in that, It also includes a data synchronization and consistency assurance module connected to the dynamic feedback control module, which is used to establish a bidirectional data synchronization channel between the stream processing engine and the big data processing framework, and ensure data consistency through a two-phase commit protocol.

7. The adaptive enterprise strategic decision-making management system according to claim 1, characterized in that, The multi-source data acquisition and integration module is also used to perform automated feature engineering in the big data processing framework. The automated feature engineering includes: extracting statistical features, time-series features, cross features and embedding features from the acquired data, and merging the extracted features.

8. The adaptive enterprise strategic decision-making management system according to claim 1, characterized in that, The visual interaction module includes: The decision interpretation unit is used to obtain decision path information and key influencing factors from the intelligent decision model module, and to display the decision path and the distribution of the key influencing factors in the decision path in a visual manner. The human-machine collaborative decision-making unit is used to display the intelligent resource configuration suggestions and their confidence levels, and to receive the decision-maker's interactive operations, using the interactive operations as feedback data.

9. The adaptive enterprise strategic decision-making management system according to claim 1, characterized in that, It also includes a self-evolution mechanism module, used for: Establish a set of key performance indicators and regularly evaluate the system's actual performance on each indicator; The actual performance is compared with the preset target value. When the difference exceeds the threshold, the model optimization process is triggered. In the optimization process, the performance of different model versions and algorithm architectures is recorded and compared through the model management toolchain, and the model with the best performance is automatically selected for deployment, thereby realizing the dynamic updating and introduction of the intelligent decision model module. Furthermore, when the model optimization process is triggered, new feature types, new neural network model structures, or new model integration methods are explored through automated machine learning technology to achieve dynamic introduction of system functions and adaptive improvement of decision performance.

10. An adaptive enterprise strategic decision-making management method, characterized in that, An adaptive enterprise strategic decision-making management system according to any one of claims 1 to 9, comprising the following steps: The multi-source data acquisition and integration process involves real-time acquisition of multi-source business data, including internal and external enterprise data, and integration and processing of the acquired data through a big data processing cluster based on a big data processing framework to output feature data. The intelligent decision-making step involves receiving the feature data and performing predictive analysis using a cascaded fusion model that includes a random forest model and a neural network model to generate market demand forecast results and intelligent resource allocation suggestions. The cascaded fusion model uses the output of the random forest model as one of the input features of the neural network model and performs weighted fusion of the results of each sub-model to generate the intelligent resource allocation suggestions. The dynamic feedback control step utilizes a stream processing engine to process data streams in real time, and based on a reinforcement learning algorithm, dynamically optimizes and adjusts the intelligent resource configuration suggestions according to the real-time business status to generate the final execution instruction. The interactive visualization process uses visual charts to show the decision-making path, key influencing factors, and real-time business metrics, and provides multi-dimensional data analysis views.

Images