Enterprise full-link intelligent management and control platform driven by multi-modal large model

Abstract

The invention relates to the technical field of data management, and discloses a multi-modal large model-driven enterprise full-link intelligent management and control platform, which comprises the following steps of: acquiring multi-modal data from different departments in an enterprise; mapping modal data collected by different departments to a unified feature vector, and obtaining modal features of the feature vector through a convolution kernel C1, a sampling pool S2 and a convolution kernel C3 in sequence; establishing a feature database of each department according to different departments and corresponding modal features; according to whether the modal features fall into feature databases of multiple departments at the same time, whether the modal features of the data of all the departments of the enterprise intersect with the modal features of other parts of data or not is judged. According to the method, periodicity and trend features are extracted by using the convolution kernel and the sampling pool, potential relevance of data of different departments can be effectively captured, data intersection is accurately identified in combination with cross validation of the feature database, and data isolation between traditional departments is broken.

Description

Technical Field

[0001] The present invention relates to the technical field of data management, and particularly to an enterprise full-link intelligent control platform driven by a multi-modal large model. Background Art

[0002] In today's digital age, the management and processing of enterprise data have become crucial. Enterprises need to be able to efficiently process a large amount of data in order to make informed decisions, improve production efficiency, and maintain a competitive edge. To meet this need, the research and development of enterprise digital transformation management methods have become very critical.

[0003] This application solves the problems of an existing patent, a method for enterprise digital transformation management (publication number CN118070342A), which relates to the field of data management. The method includes: Step 1: Obtain enterprise data and divide the enterprise data into multiple sub-enterprise data according to the organizational structure within the enterprise, and each sub-enterprise data corresponds to a minimum organizational structure within the enterprise; Step 2: Compress and encapsulate each sub-enterprise data; Step 3: Submit an enterprise data acquisition application to the minimum organizational structure within the enterprise corresponding to the sub-enterprise data corresponding to the encapsulated enterprise data, and the following problems occur: 1) It is necessary to ensure that the data coding sets between each department within the enterprise are unique and non-overlapping. If the enterprise organizational structure changes frequently, the maintenance cost of the coding system will increase significantly; 2) There is a lack of signature for the data flow process between each department within the enterprise, which makes it difficult to intuitively trace the origin of frequently changing data. Summary of the Invention

[0004] The present invention provides an enterprise full-link intelligent control platform driven by a multi-modal large model to solve the existing technical problems and solves the problems in the above background art.

[0005] To solve the above technical problems, according to one aspect of the present invention, more specifically, an enterprise full-link intelligent control platform driven by a multi-modal large model includes the following steps: S1. Collect multi-modal data from different departments within the enterprise; S2. Map the modal data collected from different departments to a unified feature vector, and sequentially obtain the modal features of the feature vector through a convolution kernel C1, a sampling pool S2, and a convolution kernel C3; S3. Establish a feature database for each department according to different departments and their corresponding modal features; S4. Determine whether there is an intersection between the modal features of the data of each department in the enterprise and the modal features of the data of other parts according to whether the modal feature falls into the feature databases of multiple departments at the same time; S5. And generate a data label judgment model by fitting based on the modal feature; S6. Calculate the difference coefficient between the modal feature and the modal features in each department's feature database through the data label judgment model; S7. Based on the difference coefficient and in combination with a set threshold, determine whether the modal features of the enterprise's various department data are uniquely present. If the modal feature is uniquely present, assign an initial department signature and loop back for feedback.

[0006] Furthermore, the multi-modal data includes structured data, text, images, and time-series data.

[0007] Furthermore, the specific steps for obtaining the modal feature in step S2 are as follows: S201. Scan the convolutional kernel C1 on the feature vector and capture the periodic feature m1 of the feature vector by calculating the weighted sum of the local area; S202. After pooling the periodic feature m1 through the sampling pool S2, extract the overall trend of the periodic feature m1; S203. Scan the convolutional kernel C3 on the overall trend to obtain the same trend feature m2 in the overall trend; S204. Establish based on the periodic feature m1 and the same trend feature m2.

[0008] Furthermore, when both the periodic feature m1 and the same trend feature m2 in the obtained modal feature exist, it indicates the existence of the feature vector.

[0009] Furthermore, according to the ratio of the proportion of the obtained trend feature m2 in the periodic feature m1 to the proportion of the trend feature m2 in the periodic feature m1 in the feature databases of other departments, determine whether the modal feature falls into the feature databases of other departments at the same time.

[0010] Furthermore, in step S5, the specific steps for fitting and generating a data label judgment model are as follows: S501. Manually set discrete difference coefficients according to the matching degree between the modal feature and the modal features in the feature databases of multiple other departments; S502. Establish a feature relationship x according to the difference coefficient and the ratio of the proportion of the trend feature m2 in the periodic feature m1 in the obtained modal feature to the proportion of the trend feature m2 in the periodic feature m1 in the feature databases of other departments; S503. Establish a feature relationship y according to the difference coefficient and the ratio of the trend feature m2 in the obtained modal feature to the trend feature m2 in the feature databases of other departments; S504. Based on the feature relationship x and the feature relationship y, a data label judgment model is fitted and established.

[0011] Furthermore, the data label judgment model calculates the difference coefficient between the modal feature and the modal features in each department's feature database. The calculation formula is as follows: ; In the formula, g represents the difference coefficient between the modal feature and the modal features in each department's feature database; a represents the ratio of the proportion of the trend feature m2 in the periodic feature m1 in the obtained modal feature to the proportion of the trend feature m2 in the periodic feature m1 in the feature databases of other departments; h represents the ratio of the trend feature m2 in the obtained modal feature to the trend feature m2 in the feature databases of other departments.

[0012] Furthermore, the enterprise full-link intelligent control platform driven by the multi-modal large model further includes a dynamic update module, which is used to regularly update the feature databases of each department according to the newly collected multi-modal data, and adjust the data label judgment model through an incremental learning algorithm to adapt to the dynamic changes of enterprise data.

[0013] Furthermore, the enterprise full-link intelligent control platform driven by the multi-modal large model further includes a cross-departmental collaboration optimization module. When it is judged in step S4 that there is an intersection of modal features, a data fusion strategy is generated based on the difference coefficient, and an automated workflow is triggered to coordinate multiple departments to execute joint tasks.

[0014] For the enterprise full-link intelligent control platform driven by the multi-modal large model provided by the present invention, compared with the prior art, the effects achieved by this method are as follows: 1. By mapping multi-modal data to a unified feature vector and using convolutional kernels and sampling pools to extract periodic and trend features, the present invention can effectively capture the potential correlations of data from different departments, and combined with the cross-validation of the feature database, accurately identify data intersections, breaking the traditional data isolation between departments.

[0015] 2. By regularly adjusting the feature database and the data label judgment model through an incremental learning algorithm, the present invention supports the rapid adaptation to newly collected data. This mechanism avoids data drift problems caused by business expansion or process adjustment, ensures the long-term effectiveness of the model, reduces the computational cost of full-scale retraining, and improves the system efficiency and forward-looking.

[0016] 3. When an intersection of modal features is detected, the present invention automatically generates a data fusion strategy based on the difference coefficient and triggers a joint task workflow, such as coordinating the sales and production departments to optimize inventory management, shortening the decision response time, reducing manual intervention errors, and achieving efficient resource allocation and cost savings.

[0017] 4. Through the calculation of the coefficient of variation and the threshold determination, the present invention can verify the uniqueness of the data modal characteristics and assign an initial department signature to the uniquely existing data. This mechanism not only strengthens the transparency of data attribution but also supports full-link traceability, providing technical guarantees for enterprise data governance and compliance management. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a schematic structural diagram of the present invention; Figure 2 is a schematic diagram of the coefficient of variation g and the ratio a of the acquired data to the database data in the present invention; Figure 3 is a schematic diagram of the coefficient of variation g and the ratio h of the acquired data to the database data in the present invention; Figure 4 is a schematic diagram of the ratio a of the database data and the ratio h of the acquired data to the database data in the present invention; Figure 5 is a schematic diagram of the eigenvector in the present invention; Figure 6 is a schematic diagram of the periodic feature m1 in the present invention; Figure 7 is a schematic diagram of the trend feature m2 in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] To make the technical solution of the present invention clearer, the following further describes the present invention in detail with reference to the accompanying drawings and specific embodiments.

[0020] Embodiment 1

[0021] As Figure 1 、 Figures 5 - 7 shown, according to one aspect of the present invention, there is provided an enterprise full-link intelligent control platform driven by a multi-modal large model, including collecting multi-modal data from different departments within the enterprise, where the multi-modal data includes structured data, text, images, and time-series data; mapping the modal data collected from different departments to a unified eigenvector. The specific steps for obtaining the unified eigenvector are as follows: The following takes the mapping formula of the text modality and details the definitions, functions, and training methods of each parameter. The parameter logic of the image and time-series modalities is similar to that of the text modality, and only the input features and independent parameters of the corresponding modalities need to be replaced. Then the mapping formula of the text modality is: ; And, K represents the number of kernel functions. It is used to control the complexity and expressive power of the model. The more kernels, the more refined the model; Denote the weight of the k-th kernel. It is used to adjust the contribution of each kernel in the final feature map and learn the importance of different kernels; t represents the original feature vector of the text modality. It is used as the input of the mapping to reflect the semantic information of the original data. We have: ; In the above formula, denotes the center point of the k-th kernel in the text feature space. It is used to define the position of the kernel and determine the response region of the kernel in the feature space; denotes the Gaussian function bandwidth of the k-th kernel. It is used to control the width of the kernel. When is larger → the kernel is narrower, the response range is small, and it captures local features. When is smaller → the kernel is wider, the response range is large, and it captures global features; denotes the Gaussian kernel function, in the form of . It is used to non-linearly map the input features to a high-dimensional space and enhance the feature separability (after processing the text data through the mapping formula of the text modality, a vector value is generated as the feature vector, as shown in Figure 5 ).

[0022] As Figures 5 - 7 shown, the modal features of this feature vector are obtained by passing through the convolutional kernel C1, the sampling pool S2, and the convolutional kernel C3 in sequence; the specific steps to obtain the modal features in this step are: S201. Scan the convolutional kernel C1 on the feature vector and capture the periodic feature m1 of this feature vector by calculating the weighted sum of the local region (obtain the periodic feature m1 of this feature vector. As Figure 6 shown, where the periodic feature in is approximately the same as the periodic feature in , then it means that this approximately same periodic feature is the periodic feature m1).

[0023] S202. After pooling the periodic feature m1 through the sampling pool S2, extract the overall trend of the periodic feature m1.

[0024] S203. Scan the convolutional kernel C3 on this overall trend to obtain the same trend feature m2 in this overall trend (the same trend feature is used to represent the curve change rate in Figure 6 . As Figure 7 shown, and have similar trend features, and as Figure 7 the x and y trend features in

[0025] S204. Based on the periodic feature m1 and the same trend feature m2 to establish.

[0026] Among them, when both the periodic feature m1 and the same-trend feature m2 in the modal features are present, it indicates the existence of the feature vector.

[0027] Embodiment 2

[0028] As Figure 1 shown, establish a feature database for each department according to different departments and their corresponding modal features; judge whether there is an intersection between the modal features of the data of each department of the enterprise and the modal features of the data of other parts according to whether the modal features fall into the feature databases of multiple departments at the same time. If there is an intersection, assign the signature of the additional department to the data with the intersection.

[0029] Among them, according to the ratio of the obtained trend feature m2 in the periodic feature m1 to the ratio of the trend feature m2 in the periodic feature m1 in the feature databases of other departments, judge whether the modal feature falls into the feature databases of other departments at the same time. By proposing to establish independent feature databases based on the modal features of each department and judging the relevance of cross-department data by analyzing whether the modal features fall into multiple databases at the same time. The core lies in detecting potential intersections between data of different departments through feature mapping and similarity analysis. For example, by comparing the ratio differences of the trend feature m2 in the periodic feature m1, the system can identify whether the data is shared by multiple departments or there is redundancy. The advantages of this method are: 1) It can accurately locate the correlation points of cross-department data and reduce the phenomenon of data islands; 2) Through the independence and cross-validation of the feature database, improve the structured level of data management; 3) Provide a basis for subsequent data label generation and collaborative optimization, and enhance the global interpretability of enterprise data.

[0030] Embodiment 3

[0031] As Figures 1 - 4 shown, fit and generate a data label judgment model based on the modal feature; the specific steps of fitting and generating the data label judgment model in this step are: S501. According to the matching degree between the modal feature and the modal features in the feature databases of other multiple departments, manually set discrete difference coefficients. Specifically: The difference coefficient g of the matching between the modal feature and the modal features in the feature databases of each department can be represented by whether the data in the sample data needs to be reviewed and approved by other departments and the total time of review and approval.

[0032] For example, when collecting data, graphics, or text from 100 enterprise departments, if the total time taken for the data, graphics, or text of a certain enterprise department to be reviewed and approved by other departments exceeds that of the data in 50 other samples, it indicates that the difference coefficient g between this modal feature and the modal features in the feature databases of each department is 50%.

[0033] S502. Establish a feature relationship x based on the ratio of this difference coefficient to the ratio of the trend feature m2 in the obtained modal feature in the periodic feature m1 to the ratio of the trend feature m2 in the periodic feature m1 in the feature databases of other departments. Specifically: Establish a mathematical model for the relationship between the difference coefficient g and the ratio a of the obtained data to the database data (as Figure 2 shown, the red dots in the figure are the distributions of the 100 collected samples), then there is: (Formula 1); In the above Formula 1, k represents an empirical constant for adjusting the sensitivity of the above model.

[0034] S503. Establish a feature relationship y based on the ratio of this difference coefficient to the ratio of the trend feature m2 in the obtained modal feature to the trend feature m2 in the feature databases of other departments. Specifically: Establish a mathematical model for the relationship between the difference coefficient g and the ratio h of the obtained data to the database data (as Figure 3 shown, the red dots in the figure are the distributions of the 100 collected samples), then there is: (Formula 2); In the above Formula 2, k represents an empirical constant for adjusting the sensitivity of the above model.

[0035] S504. Based on this feature relationship x and feature relationship y, fit and establish a data label judgment model. Specifically: Establish a mathematical model for the relationship between the difference coefficient g and the ratio a of the obtained data to the database data, and the ratio h of the obtained data to the database data (as Figure 4 shown, through which the linear relationship between the ratio a of the obtained data to the database data and the ratio h of the obtained data to the database data can be known), and combined with the feature relationships of the above Formula 1 and Formula 2, then there is: .

[0036] Example 4

[0037] Such as Figure 1 , Figures 5 - 7As shown, the difference coefficient between the modality feature and the modality features in the feature databases of each department is calculated through a data label judgment model; based on this difference coefficient and combined with a set threshold, it is determined whether the modality features of the data of each department of the enterprise are uniquely present. If the modality feature is uniquely present, an initial department signature is assigned and looped back. The data label judgment model calculates the difference coefficient between the modality feature and the modality features in the feature databases of each department, and the calculation formula is: .

[0038] In the formula, g represents the difference coefficient between the modality feature and the modality features in the feature databases of each department; a represents the ratio of the proportion of the trend feature m2 in the periodic feature m1 in the obtained modality feature to the proportion of the trend feature m2 in the periodic feature m1 in the feature databases of other departments; h represents the ratio of the trend feature m2 in the obtained modality feature to the trend feature m2 in the feature databases of other departments.

[0039] The explanations are: The ratio of the proportion of the trend feature m2 in the periodic feature m1 in the obtained modality feature to the proportion of the trend feature m2 in the periodic feature m1 in the feature databases of other departments represents: ; And there is: ; Then, .

[0040] The representation of the ratio of the trend feature m2 in the obtained modality feature to the trend feature m2 in the feature databases of other departments is: .

[0041] Examples are: Such as Figures 5 - 7 shown, in the periodic feature m1 of the trend feature m2 (such as Figure 7 the x part in ) and the trend feature m2 in Figure 7 (such as ; In the above calculation, the two 0.5 respectively represent the range of k for the x part and the range of k for the y part. The two 2 respectively represent is the range of k and when k . Then if we take , then a = 0.625. And, the ratio of the trend feature m2 in the obtained modality feature to the trend feature m2 in the feature databases of other departments takes ; It can be known from the calculation according to the above formula that the difference coefficient between the modal feature and the modal features in the feature databases of each department is . And by comparing multiple groups of implementation data, there are: Table 1 Judgment on Whether the Implementation Data and the Obtained Data Are Unique

[0042] Based on the data in Table 1 above, it can be known that when the implementation data tends to infinity, the obtained data will be divided into unique or not based on the difference coefficient g as the dividing line. For example, when , it can indicate that the obtained data exists uniquely.

[0043] Example 5

[0044] As Figure 1 shown, the enterprise full-link intelligent control platform driven by the multi-modal large model further includes a dynamic update module, which is used to regularly update the feature databases of each department according to newly collected multi-modal data, and adjust the data label judgment model through an incremental learning algorithm to adapt to the dynamic changes of enterprise data. By introducing the dynamic update module, the feature databases are regularly updated through the incremental learning algorithm and the data label judgment model is optimized. This module automatically adjusts the feature parameters according to the newly collected multi-modal data, such as the weights of the periodic feature m1 and the trend feature m2, so as to adapt to the dynamic changes of enterprise data (such as business expansion or process adjustment). Its advantages include: 1) Realize the real-time synchronization of the model and data, and avoid the model failure caused by data drift; 2) Adopt incremental learning to reduce the computational overhead of retraining and improve the system efficiency; 3) Support the continuous mining of long-term data value and ensure the forward-looking and adaptability of the control platform.

[0045] Example 6

[0046] As Figure 1 shown, the enterprise full-link intelligent control platform driven by the multi-modal large model further includes a cross-departmental collaborative optimization module. When judging whether there is an intersection between the modal features of the data of each department of the enterprise and the modal features of the data of other parts, a data fusion strategy is generated based on the difference coefficient, and an automated workflow is triggered to coordinate multiple departments to execute joint tasks. By designing the cross-departmental collaborative optimization module, when an intersection of modal features is detected, a data fusion strategy is automatically generated and the joint task workflow is triggered. For example, if there is an intersection of the trend feature m2 in the data of the sales and production departments, the system can generate an inventory optimization strategy and coordinate the two departments to synchronously execute the procurement and production scheduling plans. Its advantages are: 1) Shorten the response time of cross-departmental collaboration through data-driven workflow automation; 2) Generate precise strategies based on the coefficient of difference to reduce the error risk of manual decision-making; 3) Optimize resource allocation (such as sharing data to reduce duplicate collection) to improve the overall operational efficiency and cost-effectiveness of the enterprise.

[0047] The above-described embodiments merely represent several implementation manners of the present invention. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the patent for the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent for the present invention shall be subject to the appended claims.

Claims

1. An enterprise full-link intelligent control platform driven by a multi-modal large model, characterized in that It includes the following steps: S1. Collect multi-modal data from different departments within the enterprise; S2. Map the modal data collected from different departments to a unified feature vector, and sequentially pass through the convolution kernel C1, sampling pool S2, and convolution kernel C3 to obtain the modal features of the feature vector; S3. Establish a feature database for each department according to different departments and their corresponding modal features; S4. Determine whether there is an intersection between the modal features of the data of each department in the enterprise and the modal features of the data of other parts according to whether the modal feature falls into the feature databases of multiple departments at the same time; S5. And based on the modal feature, fit and generate a data label judgment model; S6. Calculate the difference coefficient between the modal feature and the modal features in the feature databases of each department through the data label judgment model; S7. According to the difference coefficient and combined with a set threshold, determine whether the modal features of the data of each department in the enterprise are uniquely present. If the modal feature is uniquely present, assign an initial department signature and loop back.

2. The enterprise full-link intelligent control platform driven by the multi-modal large model according to claim 1, characterized in that: The multi-modal data includes structured data, text, images, and time-series data.

3. The enterprise full-link intelligent control platform driven by the multi-modal large model according to claim 1, characterized in that: The specific steps for obtaining modal features in step S2 are as follows: S201. Scan the convolution kernel C1 on the feature vector, and capture the periodic feature m1 of the feature vector by calculating the weighted sum of the local area; S202. After pooling the periodic feature m1 through the sampling pool S2, extract the overall trend of the periodic feature m1; S203. Scan the convolution kernel C3 on the overall trend to obtain the same trend feature m2 in the overall trend; S204. Based on the periodic feature m1 and the same trend feature m2 to establish.

4. The enterprise full-link intelligent control platform driven by the multi-modal large model according to claim 3, characterized in that: When both the periodic feature m1 and the same trend feature m2 in the obtained modal features exist, it indicates the existence of the feature vector.

5. The enterprise full-link intelligent control platform driven by a multi-modal large model according to claim 3, characterized in that: According to the ratio of the proportion of the obtained trend feature m2 in the periodic feature m1 to the ratio of the proportion of the trend feature m2 in the periodic feature m1 in the feature databases of other departments, determine whether the modal feature falls into the feature databases of other departments at the same time.

6. The enterprise full-link intelligent control platform driven by a multi-modal large model according to claim 1, wherein: In step S5, the specific steps for fitting and generating a data label judgment model are as follows: S501. Manually set discrete difference coefficients according to the matching degree between the modal feature and the modal features in the feature databases of multiple other departments; S502. Establish a feature relationship x according to the difference coefficient and the ratio of the proportion of the trend feature m2 in the periodic feature m1 in the obtained modal feature to the ratio of the proportion of the trend feature m2 in the periodic feature m1 in the feature databases of other departments; S503. Establish a feature relationship y according to the difference coefficient and the ratio of the trend feature m2 in the obtained modal feature to the trend feature m2 in the feature databases of other departments; S504. Based on the feature relationship x and the feature relationship y, fit and establish a data label judgment model.

7. The enterprise full-link intelligent control platform driven by the multi-modal large model according to claim 3, wherein: The data label judgment model calculates the difference coefficient between the modal feature and the modal features in the feature databases of each department. The calculation formula is: ； Wherein, g represents the difference coefficient of the matching between the modal feature and the modal features in the feature databases of each department; a represents the ratio of the proportion of the trend feature m2 in the acquired modal feature in the periodic feature m1 to the proportion of the trend feature m2 in the periodic feature m1 in the feature databases of other departments; h represents the ratio of the trend feature m2 in the acquired modal feature to the trend feature m2 in the feature databases of other departments.

8. The enterprise full-link intelligent control platform driven by the multi-modal large model according to claim 1, characterized in that, It further includes a dynamic update module, which is used to regularly update the feature databases of each department according to the newly collected multi-modal data, and adjust the data label judgment model through an incremental learning algorithm to adapt to the dynamic changes of enterprise data.

9. The enterprise full-link intelligent control platform driven by the multi-modal large model according to claim 1, wherein, It further includes a cross-department collaborative optimization module. When it is determined in step S4 that there is an intersection of modal features, a data fusion strategy is generated based on the difference coefficient, and an automated workflow is triggered to coordinate multiple departments to execute joint tasks.

Images