Distributed data acquisition system based on multi-agent collaborative decision and construction method

Abstract

The invention belongs to the field of artificial intelligence, and discloses a distributed data acquisition system based on multi-agent collaborative decision and a construction method, and the method comprises the steps: 1, collecting heterogeneous data of production equipment, an environment sensor, a man-machine interaction terminal and an enterprise information system in real time through the deployment of a distributed data acquisition network, cleaning, de-noising and standardizing the edge computing nodes to generate a structured data stream; 2, automatically extracting features and laws of structured data through a multi-dimensional engine analysis model, and completing semantic alignment and fusion of heterogeneous data; 3, multiple agents are adopted to form empirical data according to a deterministic strategy gradient model in combination with a multi-dimensional reward function optimization model; and step 4, performing incremental learning on the empirical data to optimize the anomaly detection model, realizing multi-dimensional information visualization display of first-line decision and center command, and deploying corresponding functional agents according to the multi-dimensional information. The problems that a multi-agent system is low in control precision and low in efficiency are solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of artificial intelligence, and particularly relates to a distributed data acquisition system based on multi-agent collaborative decision and a construction method. BACKGROUND

[0002] In the daily operation of an enterprise, data related to product research and development, company governance, organization command, environmental governance, and front-line decision-making are involved. By applying intelligent agents at each node, management personnel can optimize management solutions and strategies and improve management efficiency by using multiple intelligent agents with multiple data points. Intelligent agents play a key role in various scenarios of daily management and drive the enterprise to achieve data-driven intelligent operation.

[0003] However, the multi-agent system has the following defects in actual application: first, the data integration capability is insufficient, it is difficult to handle the "data island" problem caused by massive heterogeneous data, and there is a lack of effective support of edge computing nodes, so that real-time data cleaning and standardization cannot be achieved; second, production traceability relies on a relational database, lacks a tamper-proof mechanism, and is difficult to express complex production relationships, so the data integrity and credibility are insufficient; third, the decision-making model focuses on historical data review and analysis, lacks the ability to predict production trends, and cannot meet the multi-dimensional information needs of front-line decision-making and central command; fourth, the multi-agent distributed deployment and cooperation mechanism are not optimized, the data acquisition efficiency and accuracy are limited, and it is difficult to adapt to complex production scenarios of human-machine collaboration. SUMMARY

[0004] The application provides a distributed data acquisition system based on multi-agent collaborative decision and a construction method, which solves the problems of low control precision and low efficiency of the multi-agent system.

[0005] One basic scheme provided by the application is a distributed data acquisition system based on multi-agent collaborative decision, which comprises A data acquisition module: by deploying a distributed data acquisition network, real-time acquisition of heterogeneous data of production equipment, environmental sensors, human-computer interaction terminals and enterprise information systems is performed, and cleaning, denoising and standardization processing are performed through edge computing nodes to generate structured data flow; A multi-dimensional data integration module: the characteristics and rules of structured data are automatically extracted by a multi-dimensional engine analysis model to complete semantic alignment and fusion of heterogeneous data; A collaborative decision-making module: experience data is formed by a multi-agent according to a deterministic policy gradient model combined with a multi-dimensional reward function optimization model; An intelligent decision-making module: the experience data is optimized by an incremental learning model to realize multi-dimensional information visualization display of front-line decision-making and central command, and to deploy corresponding function intelligent agents according to multi-dimensional information.

[0006] Preferably, the data acquisition module comprises a data acquisition submodule, a data preprocessing submodule, the data acquisition submodule is configured to acquire heterogeneous data; the data preprocessing submodule is configured to classify and label the acquired heterogeneous data, and perform cleaning, denoising and standardization processing on the labeled heterogeneous data, wherein the classification and labeling of the heterogeneous data comprises data type labeling, and the data type labeling comprises time series data, structured data, unstructured data, image / video data.

[0007] Further preferably, the multi-dimensional data integration module comprises a feature extraction submodule, a semantic alignment submodule, a fusion submodule, the feature extraction submodule is configured to select a corresponding feature extraction method according to the data type of the heterogeneous data to extract data features, wherein the feature extraction method comprises ResNet-50 network, CapsNet network, LSTM and TF-IDF; the semantic alignment submodule is configured to map the features of the heterogeneous data to the same semantic space by using an adversarial learning algorithm to achieve semantic alignment; the fusion submodule is configured to fuse the aligned heterogeneous data by using an attention mechanism.

[0008] Preferably, the intelligent decision module comprises an anomaly detection model optimization submodule, a visualization submodule, and a dynamic deployment submodule, the anomaly detection model optimization submodule is configured to output an anomaly probability score and an anomaly type corresponding to each heterogeneous data by using an anomaly detection model, dynamically optimize the parameters of the anomaly detection model based on experience data by using incremental learning, and generate a decision suggestion according to the anomaly probability score, the anomaly type, and a first-line decision corresponding to the heterogeneous data; the visualization submodule is configured to display the heterogeneous data, the anomaly probability score, the anomaly type, the first-line decision, and the decision suggestion; the dynamic deployment submodule is configured to automatically call or terminate an intelligent agent according to the anomaly type and the decision suggestion.

[0009] Further preferably, the decision suggestion generation method comprises constructing a strong correlation relationship among the heterogeneous data, the experience data, and an abnormal event by using a knowledge graph, and generating a decision suggestion according to the strong correlation relationship.

[0010] Preferably, the collaborative decision module comprises a multi-agent collaborative decision model and an experience data output submodule, wherein the multi-agent collaborative decision model comprises a policy layer, a coordination layer, and an execution layer, the policy layer is configured to generate a first-line decision based on a target and heterogeneous data, the coordination layer is configured to coordinate conflicts among multiple agents according to the first-line decision, the execution layer is configured to execute and monitor the first-line decision, and the experience data output submodule is configured to output experience data.

[0011] Further preferably, the weight distribution of the multi-dimensional reward function includes production efficiency, resource utilization rate, and task completion quality, the production efficiency weight is 0.4, the resource utilization rate weight is 0.35, and the task completion quality weight is 0.25.

[0012] Another basic scheme provided by the application is a construction method of a distributed data acquisition system based on multi-agent collaborative decision-making, which applies the above system and includes the following steps. Step 1: Real-time acquisition of heterogeneous data of production equipment, environmental sensors, human-computer interaction terminals and enterprise information systems is performed through deployment of a distributed data acquisition network, and edge computing nodes are used for cleaning, denoising and standardization processing to generate structured data streams. Step 2: The characteristics and rules of the structured data are automatically extracted by a multi-dimensional engine analysis model to complete semantic alignment and fusion of the heterogeneous data. Step 3: Experience data is formed by using multi-agents to optimize a model of a multi-dimensional reward function according to a deterministic policy gradient model. Step 4: The experience data is used to optimize an anomaly detection model through incremental learning, multi-dimensional information visualization display of front-line decision and central command is realized, and corresponding functional agents are deployed according to the multi-dimensional information.

[0013] The principle and advantages of the application are as follows: 1. Multi-source heterogeneous data is acquired through a distributed acquisition network, the heterogeneous data is preprocessed to form structured data, multiple data points are integrated, and recording and analysis of data of people, funds, materials, events and methods are realized, effectively solving the problems of data silos and inconsistent information in the prior art, and improving the real-time monitoring and accurate tracing ability of the production process; the multi-dimensional engine analysis model is used to extract features and fuse data of the structured data, which can automatically learn and extract features and rules in the data, greatly improving the accuracy and depth of data analysis, and realizing more intelligent decision support. 2. The collaborative decision-making module is used to coordinate conflicts between multiple agents to ensure the execution of front-line decision-making, the intelligent decision-making module is used to optimize the anomaly detection model using experience data of multiple agents, and visual display is performed, and finally the analysis structure of each model is fed back to the multiple agents to form a closed loop.

[0014] Preferably, the intelligent decision-making module is used to build strong correlations between heterogeneous data, experience data and abnormal events, and generate decision suggestions according to the strong correlations, the decision generation has strong logicality and traceability, and the accuracy of the decision is ensured.

[0015] Preferably, the multi-dimensional data integration module performs feature extraction, semantic alignment and data fusion on the data, selects a suitable feature extraction method according to the data type during feature extraction, can process multi-source data, and is suitable for industrial production and large enterprise management. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a system logic block diagram of the present application; Figure 2 is a flowchart of the present application. DETAILED DESCRIPTION

[0017] Further details are described below through specific embodiments: The specific implementation process is as follows: Embodiment 1 Referring to Figure 1 A distributed data acquisition system based on multi-agent collaborative decision-making includes A data acquisition module: by deploying a distributed data acquisition network, real-time acquisition of heterogeneous data of production equipment, environmental sensors, human-computer interaction terminals and enterprise information systems is performed, cleaning, denoising and standardization processing are performed through edge computing nodes, and structured data flow is generated; A multi-dimensional data integration module automatically extracts features and rules of structured data through a multi-dimensional engine analysis model, and completes semantic alignment and fusion of heterogeneous data; A collaborative decision-making module: experience data is formed by a multi-agent according to a deterministic policy gradient model combined with a multi-dimensional reward function optimization model; An intelligent decision-making module: experience data is optimized through incremental learning of an anomaly detection model, multi-dimensional information visualization display of first-line decision and central command is realized, and corresponding functional agents are deployed according to multi-dimensional information.

[0018] The experience data is state, action and reward in the decision-making process, the state is an environment state faced by an agent during decision-making, including: heterogeneous data features: such as real-time running parameters of production equipment, monitoring values of environmental sensors, business data of enterprise information systems, etc. System global state: cooperation state of multi-agent, resource allocation, task progress, etc.

[0019] The action is a decision-making behavior of an agent based on the current state, for example: adjusting data acquisition frequency, triggering anomaly detection model optimization, deploying or terminating a certain functional agent.

[0020] The reward is a multi-dimensional evaluation of the system feedback after performing the action, which is calculated according to a preset reward function, such as production efficiency, resource utilization, and task completion quality. Experience data is the "memory carrier" of multi-agent collaborative decision-making, which records the state, action and reward during the decision-making process, and provides the system with the ability to learn from history. This data-driven optimization mechanism enables the distributed data acquisition system to continuously evolve in complex scenarios, achieving more efficient data processing and more intelligent decision support.

[0021] Preferably, the distributed data acquisition network is composed of multiple edge computing nodes, sensor nodes and human-computer interaction terminals. The edge computing node adopts an Intel Xeon E5-2680 v4 processor and has a memory capacity of 24 GB, which is used for preliminary processing of the collected data. The sensor nodes include temperature sensors, humidity sensors, infrared sensors, pressure sensors and vibration sensors, which are used for environmental parameter monitoring and equipment operating state monitoring, respectively. Among them, the measurement range of the temperature sensor is -40℃ to 125℃, and the accuracy is ±0.5℃; the detection distance of the infrared sensor is 0.5 meters to 5 meters, which is used for human parameter monitoring. The human-computer interaction terminal is equipped with a biometric identification module, including fingerprint recognition and facial recognition functions, supporting manual behavior or identity authentication data acquisition, and the recognition accuracy reaches 99.7%.

[0022] Preferably, the data acquisition module includes a data acquisition submodule, a data preprocessing submodule, The data acquisition submodule is used for collecting heterogeneous data; specifically, the data of the production equipment includes the operating state of the equipment, such as running / stoppage / fault codes, real-time sensor readings such as temperature, vibration, current, pressure, control instructions such as PLC output signals, motor speed set values; environmental sensor data includes environmental data such as temperature, humidity, light, and noise in the workshop, energy consumption such as current, voltage, power factor, safety monitoring such as smoke concentration, infrared radiation (fire warning); human-computer interaction data includes operation records such as work order filling, button operation log (HMI interface), biometric identification such as fingerprint, facial recognition, voiceprint authentication, voice instructions: maintenance personnel voice recording; enterprise information system data includes production plans such as order information (product model, delivery date) in the ERP system, material inventory such as real-time inventory status in the WMS database, quality reports: quality inspection results (pass rate, defect type) in the MES.

[0023] The data preprocessing submodule is used for classifying and labeling the collected heterogeneous data, and performing cleaning, denoising and standardization processing according to the labeling of the heterogeneous data. The heterogeneous data classification labeling is data type labeling, and the data type labeling includes time series data, structured data, unstructured data, image / video data.

[0024] The data preprocessing submodule includes the following data processing strategies for different classification annotation types Time series data: the cleaning method adopts anomaly point detection, the denoising algorithm adopts LSTM auto-encoder reconstruction, and the normalization processing adopts Z-score normalization. Structured data: the cleaning method adopts foreign key integrity checking, the denoising algorithm adopts principal component analysis (PCA), and the normalization processing adopts One-Hot encoding. Unstructured data: the cleaning method adopts regular expression extraction of key fields, the denoising algorithm adopts TF-IDF weighted dimension reduction, and the normalization processing adopts word embedding (Word2Vec). Image / video data: the cleaning method adopts fuzzy frame elimination, the denoising algorithm adopts Gaussian filtering + edge enhancement, and the normalization processing adopts pixel normalization (0-255→0-1).

[0025] Further preferably, the classification annotation type further includes security level classification, which includes data sensitivity (such as process secret level) and access permission label (role classification), and through security level classification annotation of heterogeneous data, subsequent partition storage and management are facilitated.

[0026] Preferably, the multi-dimensional data integration module includes a feature extraction submodule, a semantic alignment submodule, and a fusion submodule, The feature extraction submodule extracts data features according to the data type of the heterogeneous data and selects a corresponding feature extraction method, which includes ResNet-50 network, CapsNet network, LSTM, and TF-IDF. Specifically, ResNet-50 network is used to process image data such as device running state images and human-computer interaction interface screenshots, and in this embodiment, the input size of the ResNet-50 network is 224x224 pixels, and the output is a 2048-dimensional feature vector; LSTM network is used to process time series data such as device running parameters and environmental monitoring data, and in this embodiment, the LSTM network adopts a bidirectional structure, the hidden layer dimension is 512, the time step is 60, and the output is a 1024-dimensional feature vector; CapsNet network is used to process structured data such as multi-dimensional data in enterprise management systems; in this embodiment, the main capsule layer of the CapsNet network contains 32 16-dimensional capsules, and the output is a 512-dimensional feature vector; TF-IDF is used to process unstructured data such as fault reasons of maintenance work orders, device log texts, and operation manuals, and the output dimension is 10,000.

[0027] The semantic alignment sub-module maps the features of the heterogeneous data to the same semantic space by using an adversarial learning algorithm to achieve semantic alignment. The adversarial learning algorithm specifically uses WGAN-GP, including a generator and a discriminator. The generator includes a three-layer fully connected network and uses a LeakyReLU activation function. The discriminator includes seamless convolution, and the loss function uses the Wasserstein distance + gradient penalty term.

[0028] The fusion sub-module fuses the aligned heterogeneous data using an attention mechanism.

[0029] Specifically, when performing fusion, the features are first converted to a uniform dimension, then the importance weight of each feature is calculated through a multi-head attention mechanism, and finally the features are weighted and fused according to the attention weight.

[0030] Preferably, the collaborative decision-making module includes a multi-agent collaborative decision-making model and an experience data output sub-module. The multi-agent collaborative decision-making model includes a policy layer, a coordination layer, and an execution layer. The policy layer generates a first-line decision based on the target and the heterogeneous data. The coordination layer is used to coordinate conflicts between multiple agents according to the first-line decision. The execution layer is used for the execution and monitoring of the first-line decision. The experience data output sub-module is used to output the experience data of the distributed data acquisition system.

[0031] Further preferably, the weight distribution of the multi-dimensional reward function includes production efficiency, resource utilization rate, and task completion quality. The production efficiency weight is 0.4, the resource utilization rate weight is 0.35, and the task completion quality weight is 0.25. Specifically, the policy layer uses an Actor network composed of four fully connected layers with hidden layer dimensions of 256, 512, 256, and 128, respectively. The activation function is ReLU, and the output layer activation function is tanh. The coordination layer uses a Critic network composed of four fully connected layers with hidden layer dimensions of 256, 512, 256, and 128, respectively. The activation function is ReLU, and the output layer does not use an activation function. During training, the target network update rate τ is set to 0.01, and the discount factor γ is set to 0.95.

[0032] In the collaborative decision-making module, the experience replay technique is used to update the model, the sampling rate is 0.18, and the replacement rate is 0.28. The experience pool capacity is set to 1 million records. Each time the data is randomly sampled from the experience pool at a rate of 18% for batch updating, and the old data in the experience pool is replaced at a rate of 28%. To improve training efficiency, a priority experience replay strategy is used. Different sampling weights are assigned to samples according to the time difference error. The α parameter is set to 0.6, and the β parameter is initially set to 0.4 and linearly increased to 1.0 during training.

[0033] Preferably, the intelligent decision module comprises an anomaly detection model optimization submodule, a visualization submodule, and a dynamic deployment submodule, The anomaly detection model optimization submodule outputs anomaly probability scores and anomaly types corresponding to each heterogeneous data through an anomaly detection model, dynamically optimizes parameters of the anomaly detection model through incremental learning based on experience data, and generates decision suggestions according to the anomaly probability scores, the anomaly types, and the first-line decisions corresponding to the heterogeneous data. The visualization submodule is configured to display the heterogeneous data, the anomaly probability scores, the anomaly types, the first-line decisions, and the decision suggestions.

[0034] The dynamic deployment submodule is configured to automatically invoke or terminate an intelligent agent according to the anomaly types and the decision suggestions.

[0035] Further preferably, the decision suggestion generation method comprises constructing strong correlation relationships among the heterogeneous data, the experience data, and the abnormal events through a knowledge graph, and generating decision suggestions according to the strong correlation relationships. The specific steps of generating decision suggestions through the knowledge graph comprise: 1) Constructing a knowledge graph A. Data modeling and knowledge graph ontology design: defining entity types, entity attributes, and entity relationships to form a knowledge graph ontology architecture according to the classified and annotated heterogeneous data, experience data, and abnormal events; B. Relationship extraction and graph filling: extracting the heterogeneous data and the experience data, mapping the extracted data to the entity types, entity attributes, and entity relationships in the knowledge graph ontology architecture, and forming a structured data set in a unified format from the mapped entity types, entity attributes, and entity relationships; C. Based on real-time heterogeneous data and historical experience data, updating the knowledge graph through iterative optimization and storing.

[0036] 2) Generating decision suggestions based on the knowledge graph according to the heterogeneous data and the experience data; a. Abnormal event association query; b. Causal reasoning and suggestion generation; c. Decision suggestion optimization and feedback.

[0037] For example, the collected data is the pressure of a chemical plant pipeline, an abnormal pressure is detected → the graph queries the historical work orders associated with “P-101” → recommends “check the sealing ring”; combined with real-time inventory (knowledge graph) to confirm that the sealing ring spare parts are available → dynamically deploy the maintenance robot.

[0038] Further preferably, during the iterative optimization of the knowledge graph, periodic supplement and correction of key events are also included, with a correction period of 10 days and a correction ratio of 0.09. The correction process includes adding newly discovered key events, updating the attribute information of existing events, adjusting the association between events, etc. The correction operation is automatically performed by the system and combined with manual review. The system automatically generates correction suggestions based on newly collected data, which are reviewed and confirmed by domain experts before being updated to the knowledge graph.

[0039] Preferably, the anomaly detection model is incrementally learned through transfer learning, with a transfer ratio of 0.55. The anomaly detection model is based on a self-encoder structure, containing an encoder and a decoder. The encoder consists of 3 fully connected layers with dimensions of 512, 256, and 128, respectively; the decoder also consists of 3 fully connected layers with dimensions of 128, 256, and 512, respectively. During the incremental learning process, 55% of the original model's parameter weights are preserved, and the remaining parameters are fine-tuned with a learning rate set to 1 / 10 of the original learning rate, i.e., 0.00003. The anomaly detection threshold is determined by the 99.7th percentile of the reconstruction error distribution of normal data, and is dynamically adjusted every week based on new data.

[0040] D3.js library is used to realize dynamic visualization of multi-dimensional information, with an update frequency of 7 minutes. The visualization interface includes five main modules: production status overview, equipment operation monitoring, abnormal event tracking, resource allocation status, and decision suggestions. The production status overview uses a dashboard and a heat map to display key performance indicators; the equipment operation monitoring uses real-time line charts to display equipment parameter trends; the abnormal event tracking uses an event timeline and a relationship graph to display abnormal events and their associated relationships; the resource allocation status uses a Sankey diagram to display resource flow; the decision suggestions use a priority list to display system-generated decision suggestions. All visualization components support interactive operations, including zooming, filtering, drilling, etc.

[0041] The system also includes a tamper-proof database that stores production process traceability data and supports trusted queries. The tamper-proof database adopts a distributed storage architecture, combined with blockchain technology to ensure data integrity and tamper resistance. Data storage uses sharding technology to improve query efficiency and system scalability. Data access control uses a role-based access control model, assigning different data access permissions based on user roles. The database supports both SQL and NoSQL query methods to meet the query needs of different scenarios. The integrity and authenticity of production traceability data are ensured through distributed storage and tamper-proof mechanisms. Distributed storage uses the IPFS (InterPlanetary File System) protocol to store data in multiple nodes, improving data availability and fault tolerance. The tamper-proof mechanism uses Merkle tree structure and digital signature technology to generate a unique hash value for each record, and regularly chain the hash value for evidence, ensuring that data cannot be tampered with. Data access control uses a role-based access control model, assigning different data access permissions based on user roles.

[0042] In this embodiment, the device operation exception scenario is taken as an example: the collected heterogeneous data is vibration sensor data (time series, frequency 5 kHz, amplitude exceeding threshold), infrared sensor detects local temperature rise of motor (120℃ vs normal 80℃±5℃), PLC fault code E-205 (bearing overheating); the sorted heterogeneous data is input into the intelligent decision module, and the knowledge graph is associated: the historical work order shows that the same type of motor caused similar vibration characteristics due to insufficient lubrication three months ago, the inventory system shows that the lubricating grease inventory is sufficient (current inventory 23 kg, threshold 5 kg), and the production strategy suggests: 1. Immediately execute: issue a shutdown instruction through the edge node (safety level P0), deploy a maintenance robot to perform lubrication operation (which takes 15±2 minutes). 2. Long-term optimization: mark the lubrication period of the device in the MES from 7 days to 5 days, and push maintenance reminders through the visualization interface.

[0043] Taking the personnel operation risk scenario as an example: The collected heterogeneous data includes: biometric detection of unauthorized personnel entering a high-risk area (face recognition confidence 85%), pressure sensor shows that device D7 is in high-pressure test mode (50MPa), HMI operation log detects abnormal button sequence (skips safety confirmation step); the relevance obtained through the knowledge graph includes the permission database showing that the person only has L1 level access permission (requires L3 level operation of this device), and the device D7 historical event record shows that the seal has failed three times due to illegal operation. The generated decision suggestion is: 1. Real-time intervention: forcibly suspend device D7 through the edge node (safety protocol P1); 2. Push alarm to security administrator (including operation screenshot and personnel ID); 3. Start voiceprint verification to require secondary authentication (if not passed within 5 minutes, trigger access lock) Embodiment 2 Referring to Figure 2 A construction method of a distributed data collection system based on multi-agent collaborative decision-making is applied to the system and includes the following steps: Step 1, by deploying a distributed data collection network, real-time collection of heterogeneous data of production equipment, environmental sensors, human-computer interaction terminals and enterprise information systems, cleaning, denoising and standardization processing by edge computing nodes to generate structured data flow; S1-1: Collecting heterogeneous data; S1-2: Classify and label the collected heterogeneous data, and clean, denoise and standardize according to the label of the heterogeneous data; the heterogeneous data classification label is a data type label, and the data type label includes time series data, structured data, unstructured data, image / video data.

[0044] Step 2, automatically extract the features and rules of structured data through multi-dimensional engine analysis model, complete semantic alignment and fusion of heterogeneous data; S2-1: According to the data type of the heterogeneous data, select the corresponding feature extraction method to extract the data features, the feature extraction method includes CNN, LSTM, TF-IDF, GNN; S2-2: Use the adversarial learning algorithm to map the features of the heterogeneous data to the same semantic space to achieve semantic alignment; S2-3: Align the heterogeneous data with the attention mechanism for fusion.

[0045] Step 3, use multi-agent to form experience data according to the deterministic policy gradient model combined with the multi-dimensional reward function optimization model; S3-1: Input the processed heterogeneous data and corresponding features into the multi-agent collaborative decision-making model for model training, and in the multi-agent collaborative decision-making model, combine the deterministic policy gradient model with the multi-dimensional reward function optimization model; S3-2: Use the trained multi-agent collaborative decision-making model in the distributed data collection system to output experience data; the experience data is the state, action and reward in the decision-making process.

[0046] Step 4, optimize the anomaly detection model through incremental learning of the experience data, realize multi-dimensional information visualization display of first-line decision and center command, and deploy corresponding function agents according to multi-dimensional information.

[0047] S4-1: Output the abnormal probability score and abnormal type of each heterogeneous data through the anomaly detection model, dynamically optimize the parameters of the anomaly detection model through incremental learning based on the experience data, and generate decision suggestions according to the abnormal probability score, abnormal type of the heterogeneous data and first-line decision. S4-2: Data transformation is performed on the heterogeneous data, abnormal probability score, abnormal type, first-line decision, and decision suggestion, and display is performed; S4-3: An agent is automatically invoked or terminated according to the abnormal type and decision suggestion.

[0048] The above is only an embodiment of the present application, and common knowledge such as specific structures and characteristics in the scheme is not described in detail here. A person skilled in the art knows all the common technical knowledge in the field of the present application before the filing date or the priority date, can know all the prior art in the field, and has the ability to apply conventional experimental means before that date. A person skilled in the art can perfect and implement the present scheme in combination with their own ability under the guidance of the present application. Some typical known structures or known methods should not be an obstacle for a person skilled in the art to implement the present application. It should be noted that, for those skilled in the art, without departing from the structure of the present application, a number of modifications and improvements can be made, which should also be considered within the protection scope of the present application. These will not affect the implementation effect and practicality of the patent. The protection scope of the present application should be subject to the content of its claims, and the specific implementation mode and the like in the specification can be used to explain the content of the claims.

Claims

1. A distributed data acquisition system based on multi-agent collaborative decision-making, characterized in that, Comprising The data acquisition module: by deploying a distributed data acquisition network, real-time acquisition of production equipment, environmental sensors, human-computer interaction terminals and enterprise information system of heterogeneous data, through the edge computing node to clean, denoising and standardization processing, generate structured data stream; Multi-dimensional data integration module, through the multi-dimensional engine analysis model automatically extracts the characteristics and rules of structured data, completes the semantic alignment and fusion of heterogeneous data; Collaborative decision module, using multi-agent according to deterministic policy gradient model combined with multi-dimensional reward function optimization model to form experience data; Intelligent decision module: the experience data is optimized by incremental learning to optimize the anomaly detection model, realizing multi-dimensional information visualization display of one-line decision and center command, and deploying corresponding function intelligent agent according to multi-dimensional information.

2. The multi-agent collaborative decision-making based distributed data acquisition system according to claim 1, characterized in that: The data acquisition module comprises a data acquisition sub-module and a data preprocessing sub-module, The data acquisition sub-module is used for collecting heterogeneous data; The data preprocessing sub-module is used for classifying and labeling the collected heterogeneous data, and cleaning, denoising and standardization processing according to the labeling of the heterogeneous data, wherein the heterogeneous data classification labeling is data type labeling, and the data type labeling includes time series data, structured data, unstructured data, image / video data.

3. The multi-agent collaborative decision-making based distributed data acquisition system according to claim 2, characterized in that: The multi-dimensional data integration module comprises a feature extraction sub-module, a semantic alignment sub-module and a fusion sub-module, The feature extraction sub-module extracts data features according to the data type of the heterogeneous data by selecting corresponding feature extraction methods, wherein the feature extraction methods include ResNet-50 network, CapsNet network, LSTM and TF-IDF; The semantic alignment sub-module maps the features of the heterogeneous data to the same semantic space by using an adversarial learning algorithm to realize semantic alignment; The fusion sub-module fuses the aligned heterogeneous data by using an attention mechanism.

4. The multi-agent collaborative decision-making based distributed data acquisition system according to claim 1, characterized in that: The intelligent decision module comprises an anomaly detection model optimization sub-module, a visualization sub-module and a dynamic deployment sub-module, The anomaly detection model optimization sub-module outputs the abnormal probability score and the abnormal type of each heterogeneous data through the anomaly detection model, dynamically optimizes the parameters of the anomaly detection model based on the experience data through incremental learning, and generates decision suggestions according to the abnormal probability score, the abnormal type of the heterogeneous data and the one-line decision; The visualization sub-module is used for displaying the heterogeneous data, the abnormal probability score, the abnormal type, the one-line decision and the decision suggestion; The dynamic deployment sub-module is used for automatically calling or terminating intelligent agents according to the abnormal type and the decision suggestion.

5. The distributed data acquisition system based on multi-agent collaborative decision-making according to claim 4, characterized in that: The decision suggestion generation method comprises constructing strong correlation between heterogeneous data, experience data and abnormal events through a knowledge graph, and generating decision suggestions according to the strong correlation.

6. The multi-agent collaborative decision-making based distributed data acquisition system according to claim 1, characterized in that: The collaborative decision module comprises a multi-agent collaborative decision model and an experience data output sub-module, wherein the multi-agent collaborative decision model comprises a policy layer, a coordination layer and an execution layer, the policy layer generates one-line decision based on target and heterogeneous data; the coordination layer is used for coordinating the conflicts between multi-agents according to one-line decision; The execution layer is used for the execution and monitoring of one-line decision, and the experience data output sub-module is used for outputting experience data.

7. The multi-agent collaborative decision-making based distributed data acquisition system according to claim 6, characterized in that: The weight distribution of the multi-dimensional reward function includes production efficiency, resource utilization rate and task completion quality, the weight of production efficiency is 0.4, the weight of resource utilization rate is 0.35, and the weight of task completion quality is 0.

25.

8. A construction method of a distributed data acquisition system based on multi-agent collaborative decision-making, applied to any one of claims 1-7, characterized in that, The method comprises the following steps: Step 1, by deploying a distributed data acquisition network, real-time acquisition of heterogeneous data of production equipment, environmental sensors, human-computer interaction terminals and enterprise information systems, cleaning, denoising and standardization processing through edge computing nodes to generate structured data flow; Step 2, automatically extracting the characteristics and rules of structured data through multi-dimensional engine analysis model, completing semantic alignment and fusion of heterogeneous data; Step 3, using multi-agent to form experience data according to the deterministic policy gradient model combined with the multi-dimensional reward function optimization model; Step 4, the experience data is optimized through incremental learning of the anomaly detection model, realizing multi-dimensional information visualization display of first-line decision and central command, and deploying corresponding function agents according to multi-dimensional information.

Images