Intelligent Data Processing System for Conference Information Based on Multimodal Data Fusion

Abstract

This invention relates to the field of data processing system technology and discloses an intelligent data processing system for conference information based on multimodal data fusion. The system includes: a transaction template generation unit that calls a model from a pre-set rule base to generate a structured data container with attribute fields; a dual-channel feature extraction unit that splits voice into business data and control command channels, extracts business data into an independent buffer, and monitors control keywords; a status control unit that unlocks the container only upon receiving a confirmation command, verifies the data, writes it to fields, and generates a transaction commit command; and a synchronous execution interface unit that performs database writes. This invention achieves physical isolation between data flow and control flow through a dual-channel architecture, uses control commands as the sole logical gate, filters non-decision-related negotiation noise at the source, and ensures the atomicity and compliance of the transformation of unstructured conference data into structured management transactions.

Description

Technical Field

[0001] This invention relates to an intelligent data processing system for conference information based on multimodal data fusion, belonging to the technical field of data processing systems. Background Technology

[0002] Currently, enterprise resource planning, project management, and customer relationship management systems form the digital foundation of business processes. They rely on standardized structured data to drive budget freezes or task distribution management actions. The key data source is often administrative or business meetings involving multiple parties. Existing technologies for data processing in meeting scenarios follow a post-processing model, treating meetings as information exchange processes. Unstructured evidence is generated through recording or transcription, and manual identification and entry into the business system are relied upon after the meeting. In scenarios such as budget changes or procurement review management, the "record first, process later" model has logical flaws. Meeting negotiations include hypothetical calculations, solution games, and final confirmation. The information flow is mixed with invalid drafts and fragmented values ​​lacking constraints. Existing single-channel speech recognition solutions improve the accuracy of text conversion, but lack the ability to perceive management business logic and cannot distinguish between the discussion process and the decision result, resulting in asynchronous time and space between decision occurrence and data effectiveness.

[0003] To address the issue of data conversion efficiency, existing technologies incorporate intelligent algorithms. For example, Chinese invention patent CN119669701A discloses an intelligent analysis and decision support system for meeting data. This system uses natural language processing to extract meeting elements and random forest and BP neural network models to predict the effectiveness of decision execution. However, such statistical probability analysis models focus on semantic generalization and trend prediction of unstructured content, neglecting the rigid requirements of data atomicity and compliance for management transactions. This approach fails to meet the stringent auditing logic of modern enterprise management systems for decision confirmation and related data entry. The system lacks physical-level logical gating of administrative instructions and cannot identify and filter hypothetical trial calculations or invalid drafts in the negotiation process at the underlying level. This results in unconfirmed intermediate data penetrating the algorithm model and mixing into the business system, making it difficult to guarantee the legal validity and logical closure of generated documents.

[0004] Therefore, how to build a self-verifying structured data generation mechanism for business logic in unstructured meeting negotiations, and transform synchronous oral negotiation into atomic management transactions with administrative effectiveness and meeting the requirements of enterprise process compliance, has become the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: a conference information intelligent data processing system based on multimodal data fusion, comprising:

[0006] The transaction template generation unit is used to respond to the meeting start signal and, based on the meeting metadata, call the management business data model from the preset rule base, construct a structured data container containing several attribute fields with administrative management business logic constraints in volatile memory, and initialize the structured data container to a locked write state.

[0007] The dual-channel feature extraction unit is used to parallel split the acquired real-time voice data stream to a logically isolated business data channel and control command channel. The business data channel is used to extract business data related to meeting management transactions from the voice data stream according to preset entity recognition rules, and continuously write the business data to a candidate data cache independent of the structured data container for updates. The control command channel is used to monitor preset control keywords related to administrative decision-making and management business process control in the voice data stream and output control commands.

[0008] The state control unit is used to establish the mapping relationship between control commands and the state flow of the structured data container. When a confirmation command representing the approval of an administrative decision is received from the control command channel, the locked write state of the structured data container is released, triggering the compliance verification of the management business rules of the current data in the candidate data cache. After the verification is passed, the data is written to the attribute field to generate a transaction commit command with management effect constraints. When a rejection command is received from the control command channel, a clearing operation is performed on the candidate data cache while maintaining the locked write state of the structured data container.

[0009] The synchronous execution interface unit is used to establish a communication connection with the external enterprise management system and execute the management business flow driver and database write operation corresponding to the transaction commit command.

[0010] Preferably, the dual-channel feature extraction unit further includes a conflict data arbitration module, used to establish an access control table between user identifiers and pre-set organizational structure data, and to assign corresponding decision weight values ​​to different user identifiers based on the business type of the attribute field; within a single filling cycle of the attribute field, if multiple candidate business data originating from different user identifiers are identified, a data competition queue is constructed in the candidate data cache area, and multiple candidate business data and their corresponding decision weight values ​​are stored in the queue; a weighted filtering logic is executed to compare the decision weight values ​​of each candidate business data in the queue; when there is a unique candidate business data whose decision weight value is higher than all other candidate business data in the queue and the difference exceeds a preset difference threshold, the candidate business data is selected as the unique valid data for filling the attribute field and the data competition queue is released.

[0011] Preferably, the synchronous execution interface unit further includes a traceability index generation module, which is used to configure a circular buffer at the multimodal data acquisition end to temporarily store real-time multimodal data streams of a preset time length; in response to the trigger signal of the transaction commit instruction generated by the status control unit, it locks the timestamp corresponding to the trigger signal and extracts meeting data segments covering a preset time window before and after the timestamp from the circular buffer; it generates a unique storage index address for the meeting data segments and encapsulates the storage index address as an immutable metadata in the transaction commit instruction; wherein, when the transaction commit instruction is called in an external enterprise management system, the meeting data segments are directly indexed and reproduced through the storage index address.

[0012] Preferably, the transaction template generation unit further includes a template dynamic switching module, which is used to perform keyword matching degree monitoring in parallel during the operation of the dual-channel feature extraction unit, and to count the frequency of non-current template keywords in the current voice data stream that do not match the preset trigger rules of the current structured data container but match other data models in the preset rule base; when the frequency of non-current template keywords exceeds the switching threshold within the preset time window, a container reload instruction is triggered; in response to the instruction, the target data model is locked based on the attributes of the non-current template keywords, a new structured data container is generated in the volatile memory, and the business data that has been filled in the original structured data container and has compatible attributes is migrated to the new structured data container, and the original structured data container is destroyed.

[0013] Preferably, when executing the weighted filtering logic, the conflict data arbitration module calculates the first... The comprehensive judgment score of each candidate business data : ,in, The score is determined by comprehensive evaluation. The basic permission weights preset based on the corresponding level in the organizational structure data according to the user identifier; This is the data accuracy coefficient for this user in the corresponding historical business data of the attribute field; The duration of the candidate service data in the voice data stream; The preset effective voice reference duration; The time decay constant is preset; the conflict data arbitration module will comprehensively determine the score. Candidate business data that is the highest and exceeds the preset difference threshold is locked as valid data.

[0014] Preferably, the dual-channel feature extraction unit further includes a condition association module, which is used to scan whether there are preset condition keywords within the time window where the business data is located while extracting business data from the voice data stream to fill the attribute field; if a condition keyword is detected, a condition object associated with the attribute field is generated, and the semantic fragment associated with the condition keyword is encapsulated in the object; in response to the existence of the condition object, the state control unit forcibly blocks the switching of the structured data container to the pre-commit state, and marks the generated transaction commit instruction as a condition suspension state. The condition suspension state requires the external enterprise management system to execute the subsequent process only after the conditions defined by the semantic fragment are met.

[0015] Preferably, the business data channel also includes a sensitive data desensitization module, which is used to retrieve a pre-set sensitive data rule base based on the security level attribute of the structured data container before writing the business data into the candidate data cache area; if the business data matches a sensitive rule, an index reference pointing to the original storage address of the business data is generated, the index reference is written into the candidate data cache area only as a placeholder, and the original business data is encrypted and stored; the status control unit only calls the index reference to restore the business data to generate a transaction commit instruction when it receives a confirmation instruction and verifies that the current operator has decryption permission.

[0016] Preferably, the control command channel includes a multimodal verification module for synchronously acquiring visual gesture data streams from the meeting venue; a confirmation command is output only when a preset control keyword is detected and a preset confirmation gesture feature consistent with the control command semantics is detected within a preset time window of the preset control keyword; the status control unit uses the visual gesture data stream as an auxiliary verification signal for the control command channel to suppress false triggering under a single voice modality.

[0017] Preferably, the status control unit further includes a timeout reset module for monitoring the data retention time in the candidate data buffer area; when the data retention time in the candidate data buffer area exceeds a preset timeout threshold, and the control command channel does not output any control commands during this period, a status reset operation is automatically triggered; the status reset operation includes rolling back the structured data container to a stable state and sending a status abnormality prompt signal to the conference control terminal to request manual intervention or re-entry.

[0018] Preferably, the system is deployed in an edge-cloud collaborative hardware architecture, including: an edge computing terminal for executing the data splitting and data extraction functions of the dual-channel feature extraction unit, and maintaining a candidate data cache locally; a cloud management server for hosting a pre-set rule base and running a transaction template generation unit and a status control unit; the edge computing terminal only uploads the data in the candidate data cache to the cloud management server through an encrypted channel to complete the generation of the transaction commit instruction when a confirmation instruction is detected.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] 1. In intelligent data processing of meeting information, a logically isolated entity payload channel and administrative signaling channel are constructed. Combined with the collaborative function of the state machine gating unit, the extraction of business data and the flow of decision-making states during the meeting negotiation process are orthogonally decoupled. The voice data stream is split in parallel. The entity payload channel is used to maintain and update the candidate data temporary storage stack dynamically, and the administrative signaling channel is used to listen for discourse markers with process control attributes. The state machine gating unit only triggers the verification and locking of the temporary storage stack data when it receives a positive confirmation signal from the administrative signaling channel, and performs a clearing operation when it receives a negative veto signal. The dual-track parallel gating mechanism simulates the separation process of draft revision and final signing in administrative decision-making from the underlying logic. It requires that only data combinations that have undergone clear procedural confirmation are transformed into atomic transaction instructions. It automatically filters hypothetical calculations, procedural drafts and redundant discussions during the negotiation process, prevents non-resolution intermediate data from polluting the external management system, and ensures the atomicity and accuracy of management data generation in unstructured negotiation scenarios. Thus, a data processing barrier adapted to the characteristics of commercial administrative decision-making is built at the bottom layer of the management system.

[0021] 2. By instantiating structured data containers in response to meeting metadata, and combining this with attribute slot data constraint rules, a reverse instantiation and adsorption mechanism for unstructured meeting streams is constructed. This abandons the passive mode of full recording and uses pre-built structured data containers as logical molds with self-verification capabilities. It guides the entity payload channel to extract only entity data that conforms to specific business logic and data type constraints. Through the state machine control unit, the attribute slot filling status is monitored and logically locked in real time. This ensures that the generated transaction instructions meet the completeness requirements of the enterprise management rule definition before the pre-commit state, and that the generated management data is compliant at the moment of generation. This eliminates the risk of loss of logical dependencies or omission of key constraints due to semantic translation loss in the traditional post-event manual transcription mode, and improves the quality and availability of management data at the generation source.

[0022] 3. In the targeted feature extraction unit, an organizational structure-based permission mapping conflict data arbitration mechanism is introduced to solve the problem of multi-source data conflicts for the same attribute slot in multi-party collaborative meeting scenarios. A dynamic permission weight table is established based on the participant identifier and the pre-set organizational structure data, and a data competition queue is built in volatile memory. When multiple candidate values ​​appear in the same slot, the state machine control unit executes weighted filtering logic to automatically lock the candidate entity data with the highest decision weight value. The enterprise's administrative authority system is directly mapped to the competition arbitration operator in the data processing process, avoiding the risk that high-level management decisions will be overwritten by subsequent low-permission discussions due to simple time-series overwrite strategies, and ensuring that the final generated transaction data follows the enterprise's administrative management level and decision validity logic. Attached Figure Description

[0023] Figure 1 This is the system logic architecture and data flow diagram of the dual-channel splitting mechanism of the present invention;

[0024] Figure 2 This is a comparison chart of processing latency and resource consumption between edge-cloud collaborative architecture and pure cloud architecture of the present invention;

[0025] Figure 3 This is a sequence diagram for the arbitration and filtering of conflicting user permission weights in this invention. Detailed Implementation

[0026] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0027] This invention provides an intelligent data processing system for meeting information based on multimodal data fusion, including a transaction template generation unit, a dual-channel feature extraction unit, a state control unit, and a synchronous execution interface unit. Logically, this system constructs a dual-channel architecture that separates administrative signaling from business payloads, coupled with a state machine-based logical gating mechanism, to transform unstructured meeting data into structured management transactions, making its data format fully adaptable to the management needs of enterprise administrative office automation. The transaction template generation unit establishes the logical baseline for data processing. Responding to the meeting start signal, this unit calls a data model from a pre-set rule base based on meeting metadata, including the meeting topic and participants. The unit instantiates a structured data container in volatile memory based on the invoked data model, containing several attribute fields with data type constraints. For example, for fixed asset procurement approval meetings, the attribute fields include budget amount, supplier code, and delivery time node. The budget amount field is bound to a numerical constraint that is less than a specific threshold. During the initialization phase, the structured data container is placed in a locked write state, at which time external data cannot be directly written to the attribute fields. The dual-channel feature extraction unit is used to realize the physical layer decoupling and parallel processing of signals. This unit splits the collected real-time voice data stream in parallel to logically isolated business data channels and control command channels.

[0028] The business data channel scans the voice data stream using pre-defined entity recognition rules. These rules, built on Conditional Random Fields or Transformer-based named entity recognition algorithms, are used to extract business data that conforms to attribute field constraints. The extracted business data is not directly written to the structured data container, but rather to a separate candidate data cache. This cache uses a circular queue structure to temporarily store the latest data for the same attribute field. There are candidate values, among which For positive integers, the candidate data buffer is constructed as a fixed-length dual-pointer ring topology based on static random access memory (SRAM), configured with atomic write pointers. With read pointer Execution based on modulus Bitwise addressing is used, and physical memory pages are set via the Memory Protection Unit (MPU). The (No-Execute) attribute blocks code injection. This buffer is mapped to a non-contiguous virtual address space segment along with the structured data container. It uses the operating system kernel-level page table isolation PTI mechanism to establish a physical firewall, ensuring that business data flows can only perform overwrite loop writes within the ring buffer before being authorized by the state machine. This physically blocks the penetration of uncontrolled data into the address segment of the structured container.

[0029] The control command channel monitors discourse markers in the voice data stream. This channel runs a keyword detection engine that matches a pre-defined control keyword library in real time. This library includes confirmation and rejection command sets. Confirmation commands contain affirmative semantics such as "decided" and "approved," while rejection commands contain negative semantics such as "not allowed" and "cancelled." When these keywords are detected and the confidence level exceeds a pre-defined threshold, the channel outputs the corresponding control command. The state control unit establishes a mapping between control commands and the state transitions of the structured data container. In the default locked write state, this unit only responds to signals from the control command channel. When a confirmation command is received, the state control unit performs the following timing actions: unlocking the structured data container's locked write state and freezing the candidate data buffer. The state unlocking action is triggered by a hardware-level comparison and exchange (CAS) atomic command, which modifies the read / write permission bits of the corresponding memory page table entry (PTE) of the structured container in real time. By default, it is read-only. State flipped to writable The system synchronously inserts MemoryBarrier instructions to refresh the CPU pipeline and ensure multi-core cache consistency. It then initiates a direct memory access (DMA) controller to establish a high-speed data transfer channel from the cache area to the container fields. Immediately after the data transfer is complete, the page table permissions are reset to a locked state, achieving precise gating of the time window. This gating logic maps the authority confirmation and approval actions in administrative management to the digital space. Next, it extracts the current data from the candidate data cache and performs compliance verification based on the pre-defined data constraint rules of the structured data container. If the verification passes, the data is written to the corresponding attribute field, and the structured data container's state is switched to the pre-commit state, generating a transaction commit instruction containing the complete business context. If the verification fails, an exception log is generated, and the structured data container remains in a pending state.

[0030] Upon receiving a veto instruction, the state control unit sends a reset signal to the candidate data cache, clearing all currently cached data and maintaining the locked write state of the structured data container. For data conflicts in multi-party collaboration scenarios, the dual-channel feature extraction unit also includes a conflict data arbitration module. During the meeting initialization phase, this module loads the user identifiers of the participants, generates a permission mapping table based on organizational structure data, and assigns decision weight values ​​for different business type fields to different user identifiers. Within a single filling cycle of an attribute field, if the business data channel identifies multiple inconsistent candidate business data originating from different user identifiers, the system constructs a data competition queue in the candidate data cache, recording each candidate business data and its corresponding decision weight value. This module executes weighted filtering logic, calculating the first [value] according to the following formula. The comprehensive judgment score of each candidate business data : ,in, The score is determined by comprehensive evaluation. This refers to the pre-defined basic permission weights based on user identifiers at corresponding levels in the organizational structure data, with a value range of [missing value]. to ; This is the data accuracy coefficient for this user in the corresponding historical business data of the attribute field; The duration of the candidate service data in the voice data stream, in seconds; The preset effective voice reference duration; The time decay constant is preset. After calculation, the module compares the comprehensive judgment scores of each candidate business data in the data competition queue. When there is a unique candidate business data whose comprehensive judgment score is higher than all other candidate business data in the queue and the difference exceeds the preset difference threshold, the module selects the candidate business data as the only valid data to fill the attribute field and releases the data competition queue. The synchronous execution interface unit is used to realize the audit traceability and final submission of data. This unit maintains a circular buffer at the multimodal data acquisition end to temporarily store real-time multimodal data streams of preset time length. The circular buffer follows the first-in-first-out principle.

[0031] When the state control unit generates a transaction commit instruction, i.e., the triggering time... The synchronous execution interface unit locks the timestamp and extracts the time window from the circular buffer. The meeting data fragment, in which, The preset duration before the trigger time. For a preset duration after the trigger time, this unit generates a unique storage index address for the extracted meeting data segment and encapsulates this storage index address as immutable metadata in the transaction commit instruction. The external enterprise management system indexes and reproduces the meeting data segment by calling this storage index address. In addition, the transaction template generation unit also includes a template dynamic switching module. During system operation, this module counts in parallel the frequency of non-current template keywords in the current voice data stream that do not match the trigger rule of the current structured data container but match other data models in the preset rule base. When the frequency of non-current template keywords exceeds the switching threshold within a preset time window, this module triggers a container reload instruction. In response to this instruction, the system locks the target data model based on the attributes of the non-current template keywords, generates a new structured data container in volatile memory, migrates the already filled and attribute-compatible business data in the original structured data container to the new structured data container, and destroys the original structured data container. Finally, the synchronization execution interface unit establishes a communication connection with the external enterprise management system. This unit receives the generated transaction commit instruction and executes the database write operation corresponding to the instruction, completing the synchronization of structured data to the external system.

[0032] Example 1: In a scenario involving the emergency expansion of IT infrastructure in a large multinational corporation, budget approvals of up to five million and cross-departmental collaboration are required. The decision-making process typically occurs in a real-time voice negotiation environment involving multiple parties. In this scenario, the project manager, technical director, and finance director need to reach a consensus on the change plan within a short period, ensuring that key elements such as verbally confirmed additions of 200 servers, an additional budget of 1.5 million, and Q3 delivery are seamlessly transformed into rigid constraints for subsequent execution. The objective challenge of this scenario is that unstructured voice negotiation streams are highly volatile and ambiguous, often accompanied by implicit premises such as financial approval, leading to a temporal and spatial separation between business data and the decision-making state. Traditional post-event manual processing cannot lock the logical loop at the moment of decision-making in real time, resulting in a disconnect between purchase orders and the original decision intent. The system provided in this embodiment of the invention, after accessing this meeting scenario, has a transaction template generation unit responding to the meeting start signal. The system instantiates a structured data container for IT infrastructure changes in the background. This structured data container has pre-set attribute fields such as change amount, number of devices, and delivery deadline, and is bound with a logical constraint that the limit for a single change is less than 2 million yuan. During the negotiation process, when the technical director suggests purchasing 200 high-performance servers, and the finance director responds by approving an additional budget of 1.5 million yuan, with delivery before September 30, the dual-channel feature extraction unit performs parallel processing. The business data channel uses entity recognition rules to extract entity data such as 200 units, 1.5 million yuan, and September 30 from the voice data stream and continuously updates it to the candidate data cache. At this time, the structured data container remains locked to prevent the intrusion of ineffective data. At the same time, the control command channel monitors the voice stream in real time. When it recognizes the finance director's approval statement, which has clear administrative effect, and the confidence level exceeds the preset threshold, it outputs a confirmation command.

[0033] Upon receiving the confirmation command, the status control unit executes a logic gating operation. This unit unlocks the structured data container and triggers a compliance check on the current data in the candidate data cache. The system automatically compares the extracted budget value of 1.5 million with the pre-set limit constraints within the container. After the check passes, the business data in the candidate data cache is written to the corresponding attribute fields in one go, and the status of the structured data container is switched to the pre-commit state. This process directly drives the collapse of the data status through administrative signaling, eliminating the time lag between business negotiation and data effectiveness. Secondly, the interface unit locks the timestamp of the decision trigger moment, extracts the original voice and video clips covering the moment of the decision from the underlying circular buffer, generates a unique storage index address, and encapsulates it into the transaction commit command. Finally, the system injects a change request form containing complete business data and traceability index into the external ERP system.

[0034] Example 2: To verify the effectiveness and stability of the system proposed in this invention under complex business negotiation environments, this example constructs an experimental verification platform to simulate a typical multi-party collaborative budget review meeting scenario, including frequent topic switching, overlapping speakers, interference from non-decisive discussions, and complex interactive behaviors such as decision rollback. The platform quantitatively evaluates the system's ability to accurately identify administrative signals and precisely extract, lock, and generate compliant structured transaction instructions from mixed negotiation data when faced with high-noise and unstructured voice streams. This demonstrates the engineering practicality of the dual-channel orthogonal processing architecture and state machine logic gating mechanism of this invention. The experimental platform is configured with a multimodal data acquisition environment conforming to commercial meeting standards. The voice acquisition end uses a ring microphone array with a sampling rate set to 48kHz. To ensure sound source localization and separation capabilities in multi-person reverberation environments, the system's core processing unit is deployed on an edge server equipped with an NVIDIA T4 inference accelerator card. This unit runs the transaction template generation unit, dual-channel feature extraction unit, and state control unit of this invention. To simulate real business rule constraints, an annual budget adjustment model is loaded into the pre-set rule base. This model sets three key attribute fields: adjustment subject, adjustment amount, and effective date. The adjustment amount is also bound to a single transaction limit of no more than 500,000 yuan and must be a multiple of 1000. To introduce uncertainties from real-world engineering scenarios, office environmental noise with a signal-to-noise ratio of 15dB, including keyboard typing and distant conversations, is superimposed in the background during the experiment to test the system's noise immunity.

[0035] The experiment involved a 15-minute simulated negotiation script where three participants, acting as a project manager, finance director, and technical expert, discussed the server expansion budget. Initially, the project manager proposed a vague quote of approximately 600,000 or 550,000, while the technical expert added a desired delivery date of next Friday. At this point, although the business data channel identified candidate entities such as 600,000, 550,000, and next Friday, and pushed them into the candidate data buffer, the state control unit kept the structured data container locked because the control command channel did not detect any definitive administrative signals, effectively preventing the writing of these procedural draft data. As the discussion deepened, the finance director clearly stated... State: Based on compliance requirements, 600,000 exceeds the limit, so we set it at 480,000, effective on the 1st of next month. After saying this, a preset gesture confirmation action is executed (simulating multimodal verification). At this moment, the control command channel captures the setting of this utterance mark and visual confirmation signal, outputs a confirmation command, and the state control unit immediately triggers logic gating to extract and verify the latest 480,000 and the 1st of next month in the buffer. Since 480,000 meets the constraint conditions of being less than 500,000 and being an integer multiple of 1,000, the system verification passes, the data is quickly solidified into the attribute field, and a transaction commit command is generated. To intuitively demonstrate the processing efficiency of this invention under complex interference, Table 1 lists the key data nodes and their processing results in the experimental process.

[0036] Table 1: Key Negotiation Events and System Status Response Table

[0037]

[0038] Referring to Table 1, the experimental data shows the deterministic response logic of the system at different negotiation stages. Of particular note is the decision rollback event that occurred at 12:10. When the finance manager issued a void instruction, the system not only did not generate an incorrect transaction document, but also responded to the rejection instruction and performed a status rollback operation.

[0039] Example 3: This example combines Figures 1 to 3 Describe the intelligent data processing system for conference information based on multimodal data fusion, such as... Figure 1 As shown, the logical architecture begins with the input of meeting metadata, including the meeting topic and attendee list. The transaction template generation unit calls the model based on the data model and constraint rules in the pre-set rule base to generate a structured data container containing attribute fields and initially in a locked write state. At the same time, unstructured real-time voice and multimodal data streams are input into the system and then split by the dual-channel feature extraction unit. The business data channel is responsible for extracting business entity data and writing it into a candidate data cache with temporary storage and isolation functions, while the control instruction channel monitors control keywords in parallel to output confirmation or rejection instructions. The state control unit, as the core logical hub, performs logical gating, compliance verification, and instruction generation operations after receiving the data to be verified. Only after the instruction is confirmed does it unlock the structured data container and write the attribute fields. Finally, it switches the state to the pre-commit state and generates a transaction commit instruction, driving the synchronous execution interface unit to perform database write operations, and finally synchronizes the structured data to external enterprise management systems such as ERP, CRM, or PM systems.

[0040] like Figure 2 As shown, the horizontal axis represents the number of concurrent users from 10 to 500, the left vertical axis represents the processing latency in milliseconds, and the right vertical axis represents the resource utilization rate in percentage. The solid line in the graph shows that the edge-cloud collaborative data processing latency increases gradually with the number of users, lower than the pure cloud data processing latency shown by the dashed line. This advantage is even more pronounced in high-concurrency scenarios with 500 users. Similarly, the dotted line in the graph shows that the edge-cloud collaborative resource utilization rate is consistently lower than the pure cloud resource utilization rate shown by the long dashed line. Figure 3As shown in the sequence diagram, this diagram details the workflow of the conflict data arbitration module, involving the interaction sequence between user A (high privilege) and user B (low privilege). When user A inputs a budget of 480,000 via voice, the business data channel extracts this candidate data of 480,000, and the arbitration module records user A's weight value. When user B inputs a budget of 550,000 via voice, the channel extracts candidate data of 550,000 again and records user B's weight value. At this time, the conflict data arbitration module constructs a data competition queue, calculates the comprehensive judgment score of each candidate data, and executes weighted filtering logic. Since user A has higher privileges, the system determines that there is a unique high score that exceeds the threshold, thus selecting the valid data as 480,000 and releasing the data competition queue. Finally, the candidate data buffer holds this valid data and waits for the confirmation instruction from the status control unit. If a unique valid data cannot be determined, the system logic switches to waiting for more input.

[0041] Example 4: To address the issue of decision weight allocation in multi-party conflict scenarios and the lack of adaptive logic for dynamic template switching during long-term operation, this example provides a detailed implementation procedure for a dynamic weight calculation mechanism and template hot reload triggering logic based on organizational structure topology. It transforms the general description of conflict arbitration and template switching into a reproducible algorithm flow with clear mathematical definitions and engineering parameters, specifically addressing the decision weight values ​​in the conflict data arbitration module. The system does not rely on empirical assignment to determine the organizational structure; instead, it uses a topological distance algorithm based on the organizational structure graph for real-time calculation. The system parses the enterprise's organizational structure into a directed tree. ,in A set of nodes, representing departments or individuals; Let the set of edges represent reporting relationships, and define the node where the meeting initiator is located as the root node. The node where the current speaker is located is System computing nodes to the root node path distance and nodes Hierarchical depth in a tree .

[0042] Based on this topology, the decision weight values The computational logic is defined as follows: ,in, The administrative distance between users and the decision-making center, i.e., the meeting initiator; Based on the administrative level of the user, The highest level; This represents the maximum depth of the organizational structure tree. The preset distance attenuation coefficient, for example, is set to a value of This algorithm is used to adjust the degree to which administrative distance weakens the weights. It ensures that users who are closer to the decision-making center and have higher administrative levels have higher basic priority in the competition queue for their generated candidate data, thus achieving an objective mapping of the power and responsibility system to data weights. This is applied to the switching threshold in the template dynamic switching module. The system's setting and triggering logic incorporates a confidence accumulation mechanism based on a sliding time window to eliminate transient noise interference. The system maintenance length is... A sliding window that displays real-time statistics on the hit frequency of keywords not in the current template within the window. In this embodiment Values Seconds, switch the judgment logic to calculate keyword density in the window. : At the same time, the system introduces an inertial damping factor. To prevent frequent jittering, the final triggering condition is formalized as follows: ,in, For example, the preset keyword density threshold. times / second; for Continue to exceed The stable duration; For the minimum stability time constraint, for example The system generates a container reload instruction only when the keyword density remains above the threshold for a sufficiently long period of time.

[0043] When performing container reloading, the system uses a fuzzy matching algorithm based on attribute names for data migration, for the original container populated fields The system calculates its relationship with the new container. Each field in Name similarity In this embodiment, Levenshtein distance is used to calculate similarity. The value is greater than the preset matching threshold, and the field data types and When the data type constraints are compatible, the system will automatically... Value migration to Otherwise The data is archived to the historical log. In this embodiment, the matching threshold is set to [value]. This procedure ensures the lossless inheritance of key common data during business model switching, while avoiding data pollution caused by type mismatch.

[0044] Example 5: To address the black-box problem of initial configuration in complex industrial network environments and ensure stable spatiotemporal synchronization between the data acquisition module and heterogeneous edge devices from the initial deployment stage, this example provides a standardized pre-deployment calibration procedure. This procedure eliminates the impact of inter-device clock drift and network jitter on the alignment accuracy of multimodal data by defining clearly defined signal injection and delay measurement steps. When the system first connects to the enterprise intranet or when the network topology changes, a spatiotemporal reference calibration process is executed, and the interface unit synchronously broadcasts a set of timestamps to all registered acquisition terminals. The reference probe packet is received by each acquisition terminal, which records the local reception time the instant it receives the probe packet. and send back containing Local processing latency Upon receiving the response packets, the core processing unit collects all response packets and, based on the principles of the network time protocol, calculates the clock offset of each terminal relative to the central node. Round-trip delay Clock skew The computational logic is defined as follows: ,in, The time it takes for the terminal to send a response packet. The time it takes for the central node to receive the response packet, based on the calculated... The system automatically generates a clock compensation configuration table for each terminal and sends it to the edge side to perform local clock correction. When the absolute value of the remaining clock deviation of all terminals is... All are less than the preset synchronization threshold, for example The system determines that calibration is complete and allows the start of the formal data acquisition task only after milliseconds.

[0045] Secondly, the system has a pre-set sensitivity-encryption policy mapping table, which divides data fields into four levels: public, internal, confidential, and top secret. Before processing any business data, the business data channel uses regular expression matching and keyword scanning technology to calculate the sensitivity score of the data content. ,based on The system automatically invokes the corresponding encryption configuration within the specified interval. The instruction is classified as confidential. If specific amounts or contract terms are involved, the system will encrypt the data payload using the AES-256 algorithm and generate a temporary session key valid only for the duration of this session. The execution logic of encryption operations follows the strict structure as follows: ,in, The original plaintext data, This is a hash digest of the plaintext, used for integrity verification. For XOR operation, For encryption functions, The final generated ciphertext, the generated ciphertext The encrypted metadata (excluding the key) is written to the candidate data cache, while the key... Secure hosting is then provided through an independent key distribution center.

[0046] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A conference information intelligent data processing system based on multimodal data fusion, characterized in that, include: The transaction template generation unit is used to respond to the meeting start signal and, based on the meeting metadata, call the management business data model from the preset rule base, construct a structured data container containing several attribute fields with administrative management business logic constraints in volatile memory, and initialize the structured data container to a locked write state. A dual-channel feature extraction unit is used to split the acquired real-time voice data stream in parallel to a logically isolated business data channel and control command channel; The business data channel is used to extract business data related to meeting management from the voice data stream according to preset entity recognition rules, and continuously write the business data to the candidate data cache area, which is independent of the structured data container, to perform updates; the control command channel is used to monitor preset control keywords related to administrative decision-making and management business process control in the voice data stream and output control commands. The state control unit is used to establish the mapping relationship between control commands and the state transitions of structured data containers; When a confirmation instruction representing the approval of an administrative decision is received from the control instruction channel, the locked write state of the structured data container is released, triggering a compliance check of the management business rules for the current data in the candidate data cache. After the check passes, the data is written to the attribute field to generate a transaction commit instruction with management binding force. When a rejection instruction is received from the control instruction channel, a clearing operation is performed on the candidate data cache while maintaining the locked write state of the structured data container. The synchronous execution interface unit is used to establish a communication connection with the external enterprise management system and execute the management business flow driver and database write operation corresponding to the transaction commit instruction; In addition, the synchronous execution interface unit also includes a traceability index generation module, which is used to configure a circular buffer at the multimodal data acquisition end to temporarily store real-time multimodal data streams of a preset time length; In response to the trigger signal of the transaction commit instruction generated by the status control unit, the timestamp corresponding to the trigger signal is locked, and the meeting data segment covering the preset time window before and after the timestamp is extracted from the circular buffer; a unique storage index address is generated for the meeting data segment, and the storage index address is encapsulated as an immutable metadata in the transaction commit instruction; when the transaction commit instruction is called in the external enterprise management system, the meeting data segment is directly indexed and reproduced through the storage index address.

2. The intelligent data processing system for conference information based on multimodal data fusion according to claim 1, characterized in that, The dual-channel feature extraction unit also includes a conflict data arbitration module, which is used to establish an access control table between user identifiers and pre-set organizational structure data, and to assign corresponding decision weight values ​​to different user identifiers based on the business type of the attribute field. Within a single filling cycle of the attribute field, if multiple candidate business data originating from different user identifiers are identified, a data competition queue is constructed in the candidate data cache, and multiple candidate business data and their corresponding decision weight values ​​are stored in the queue; weighted filtering logic is executed to compare the decision weight values ​​of each candidate business data in the queue; when there is a unique candidate business data whose decision weight value is higher than all other candidate business data in the queue and the difference exceeds a preset difference threshold, the candidate business data is selected as the unique valid data for filling the attribute field and the data competition queue is released.

3. The intelligent data processing system for conference information based on multimodal data fusion according to claim 1, characterized in that, The transaction template generation unit also includes a template dynamic switching module, which is used to perform keyword matching degree monitoring in parallel during the operation of the dual-channel feature extraction unit. It counts the frequency of non-current template keywords in the current voice data stream that do not match the preset trigger rules of the current structured data container but match other data models in the preset rule base. When the frequency of non-current template keywords exceeds the switching threshold within the preset time window, a container reload instruction is triggered. In response to the instruction, the target data model is locked based on the attributes of the non-current template keywords, a new structured data container is generated in the volatile memory, and the business data that has been filled in the original structured data container and has compatible attributes is migrated to the new structured data container, and the original structured data container is destroyed.

4. The intelligent data processing system for conference information based on multimodal data fusion according to claim 2, characterized in that, When executing the weighted filtering logic, the conflict data arbitration module calculates the first value according to the following formula. The comprehensive judgment score of each candidate business data : ,in, The score is determined by comprehensive evaluation. The basic permission weights preset based on the corresponding level in the organizational structure data according to the user identifier; This is the data accuracy coefficient for this user in the corresponding historical business data of the attribute field; The duration of the candidate service data in the voice data stream; The preset effective voice reference duration; The time decay constant is preset; the conflict data arbitration module will comprehensively determine the score. Candidate business data that is the highest and exceeds the preset difference threshold is locked as valid data.

5. The intelligent data processing system for conference information based on multimodal data fusion according to claim 1, characterized in that, The dual-channel feature extraction unit also includes a condition association module, which is used to scan whether there are preset condition keywords in the time window where the business data is located while extracting business data from the voice data stream to fill the attribute field. If a condition keyword is detected, a condition object associated with that attribute field is generated, and the semantic fragment associated with the condition keyword is encapsulated in the object. In response to the existence of the condition object, the state control unit forcibly blocks the switching of the structured data container to the pre-commit state and marks the generated transaction commit instruction as a condition pending state. The condition pending state requires the external enterprise management system to execute the subsequent process only after the conditions defined by the semantic fragment are met.

6. The intelligent data processing system for conference information based on multimodal data fusion according to claim 1, characterized in that, The business data channel also includes a sensitive data desensitization module, which retrieves a pre-defined sensitive data rule base based on the security level attribute of the structured data container before writing business data into the candidate data cache. If the business data matches a sensitive rule, an index reference pointing to the original storage address of the business data is generated. The index reference is written to the candidate data cache only as a placeholder, and the original business data is encrypted and stored. The status control unit only calls the index reference to restore the business data to generate a transaction commit instruction when it receives a confirmation instruction and verifies that the current operator has decryption permission.

7. The intelligent data processing system for conference information based on multimodal data fusion according to claim 1, characterized in that, The control command channel includes a multimodal verification module, which is used to synchronously collect visual gesture data streams from the meeting venue. When a preset control keyword is detected and a preset confirmation gesture feature that is semantically consistent with the control command is detected within the preset time window of the preset control keyword, a confirmation command is output. The status control unit uses the visual gesture data stream as an auxiliary verification signal for the control command channel to suppress false triggering under a single voice modality.

8. The intelligent data processing system for conference information based on multimodal data fusion according to claim 1, characterized in that, The status control unit also includes a timeout reset module for monitoring the data retention time in the candidate data buffer. When the data retention time in the candidate data buffer exceeds a preset timeout threshold and the control command channel does not output any control commands during this period, a status reset operation is automatically triggered. The status reset operation includes rolling back the structured data container to a stable state and sending a status abnormality prompt signal to the conference control terminal to request manual intervention or re-entry.

9. The intelligent data processing system for conference information based on multimodal data fusion according to claim 1, characterized in that, The system is deployed in an edge-cloud collaborative hardware architecture, including: an edge computing terminal, which performs data splitting and data extraction functions of the dual-channel feature extraction unit and maintains a candidate data cache locally; a cloud management server, which hosts a pre-set rule base and runs a transaction template generation unit and a status control unit; the edge computing terminal only uploads the data in the candidate data cache to the cloud management server through an encrypted channel to complete the generation of the transaction commit instruction when it detects a confirmation instruction.

Images