BI data analysis method and system based on AI model

Abstract

The invention relates to a BI data analysis method and system based on an AI model, and the method comprises the steps: receiving an analysis request containing structured data and a natural language text, and carrying out the coding of the analysis request to obtain a data feature vector and a text feature vector; sub-features of different levels are extracted from the encoder multilayer network, hierarchical interactive calculation is carried out through a cross-modal attention mechanism, and aligned hierarchical features are generated; aggregating the three features, inputting the aggregated features into a dynamic weight distribution network, performing weighted fusion based on the learned convex combination parameters, and generating multi-modal fusion features in combination with residual compensation; and finally, automatically outputting a structured analysis report through the report generation model. Through hierarchical feature alignment and a dynamic fusion mechanism, the problems of semantic understanding deviation and fusion strategy stiffness existing in a traditional BI system are solved, automatic data analysis is achieved, and the accuracy of data analysis is remarkably improved.

Description

Technical Field

[0001] This application relates to the field of data analysis technology, and in particular to a BI data analysis method and system based on an AI model. Background Technology

[0002] With the acceleration of enterprise digitalization, business intelligence (BI) systems have become core tools for corporate decision-making. Traditional BI data analysis systems mainly rely on analysts and other business personnel to pre-define requirements, followed by data modeling, report development, and deployment by the development team. This process is highly dependent on the professional experience of personnel and is slow to respond to business changes, making it difficult to meet the urgent market demands for timeliness and flexibility in data analysis.

[0003] Currently, existing technologies employ an improved AI-based BI data analysis technique. This technique allows users to input queries containing natural language and utilizes pre-trained models to extract data features and text features separately. Typically, simple feature concatenation or weighted summation is used to integrate text feature vectors with data feature vectors. This fusion approach only performs a coarse-grained combination at the top-level semantic level, ignoring the inherent connections and differences between the two modalities at multiple levels and fine-grained semantics, thus failing to achieve deep semantic understanding and synergy.

[0004] The aforementioned technologies suffer from a lack of fine-grained cross-modal semantic alignment, making it difficult for the models to accurately parse the specific data dimensions or patterns referred to in the user's text description. This results in a discrepancy between the generated reports and the user's true intentions, leading to insufficient accuracy in the analysis.

[0005] Based on this, this application provides a BI data analysis method and system based on an AI model. Summary of the Invention

[0006] To address the problem that the lack of fine-grained cross-modal semantic alignment makes it difficult for models to accurately parse specific data dimensions or patterns referred to in user text descriptions, resulting in deviations between the generated reports and the user's true intentions and insufficient accuracy of the analysis, this application provides a BI data analysis method and system based on an AI model.

[0007] Firstly, this application provides a BI data analysis method based on an AI model, employing the following technical solution: including: The system receives an analysis request input by a user, the analysis request including structured data and natural language text. It uses a data feature encoder to encode the structured data to obtain a data feature vector, and uses a text feature encoder to encode the natural language text to obtain a text feature vector. Based on the data feature vector and the text feature vector, data sub-features and text sub-features at different levels are extracted from multiple network layers of the data feature encoder and the text feature encoder, respectively. Through a cross-modal attention mechanism, the data sub-features and text sub-features at the same level or adjacent levels are interactively calculated to generate aligned hierarchical features. The data feature vector, the text feature vector, and the hierarchical features are aggregated, and the contribution weights of the data feature vector, the text feature vector, and the hierarchical features in the fusion feature are dynamically allocated through a dynamic weight allocation network. Based on the contribution weights, weighted fusion is performed to generate a multimodal fusion feature that can simultaneously reflect data patterns and text semantics. The multimodal fusion features are input into the report generation model to automatically generate a structured analysis report containing data insights.

[0008] Preferably, the step of receiving the analysis request input by the user includes structured data and natural language text. A data feature encoder is used to encode the structured data to obtain a data feature vector, and a text feature encoder is used to encode the natural language text to obtain a text feature vector, including: The received structured data is preprocessed and converted into a structured data tensor that the model can process; simultaneously, the natural language text is segmented and encoded, and converted into a text tag sequence. The structured data tensor is input to the data feature encoder, which is a sequence model or graph neural network based on Transformer. The input data is encoded through its multi-layer network, and the last layer of the network outputs the global data feature vector. The text tag sequence is input into the text feature encoder, which is a pre-trained language model that encodes the input sequence through its multi-layer network, and finally outputs the global text feature vector by the last layer of the network. During the forward computation of the data feature encoder and the text feature encoder, intermediate representations are extracted synchronously from the i-th network layer of the data feature encoder as the data sub-features, and intermediate representations are extracted from the i-th network layer of the text feature encoder as the text sub-features; where i is an integer greater than or equal to 1 and less than the total number of network layers.

[0009] Preferably, the step of extracting data sub-features and text sub-features at different levels from multiple network layers of the data feature encoder and text feature encoder based on the data feature vector and the text feature vector, respectively, and performing interactive calculations on the data sub-features and text sub-features at the same level or adjacent levels through a cross-modal attention mechanism to generate aligned hierarchical features includes: For the i-th network layer of the encoder, the extracted data sub-features and text sub-features are respectively mapped to a shared d-dimensional semantic space through independent linear transformation layers to obtain mapped data sub-features and mapped text sub-features of a unified dimension. Based on the mapped data sub-features and mapped text sub-features, bidirectional cross-modal attention calculation is performed. Using the mapped text sub-features as the query and the mapped data sub-features as the key and value, text-to-data attention is calculated to generate text-guided data augmentation features. Using the mapped data sub-features as the query and the mapped text sub-features as the key and value, data-to-text attention is calculated to generate data-guided text augmentation features. The data augmentation features are combined with the text augmentation features to generate the aligned hierarchical features corresponding to the i-th network layer.

[0010] Preferably, the step of combining the data augmentation features with the text augmentation features to generate the aligned hierarchical features corresponding to the i-th network layer includes: For the i-th network layer, the data augmentation feature and the text augmentation feature are concatenated to form the primary fusion feature of the corresponding network layer; All the primary fusion features from all network layers are spliced ​​together to form a comprehensive multi-level fusion representation; The multi-level fusion representation is input into a multilayer perceptron (MLP) for nonlinear transformation and dimensionality reduction, and semantically guided multi-level embedding features are output to obtain the hierarchical features.

[0011] Preferably, the step of aggregating the data feature vector, the text feature vector, and the hierarchical features, and dynamically allocating the contribution weights of the data feature vector, the text feature vector, and the hierarchical features in the fused features through a dynamic weight allocation network, and performing weighted fusion based on the contribution weights to generate a multimodal fused feature that can simultaneously reflect data patterns and text semantics, includes: The data feature vector, the text feature vector, and the hierarchical features are each passed through an independent linear projection layer, and the projected features are then concatenated to form aggregated features. The aggregated features are input into a dynamic weight allocation network, which includes a feedforward layer and a normalization function, and outputs a set of dynamically learned convex combination parameters, including a first weight corresponding to the data feature vector, a second weight corresponding to the text feature vector, and a third weight corresponding to the hierarchical features, and the sum of the first weight, the second weight, and the third weight is 1. Based on the first weight, the second weight, and the third weight, the projected data feature vector, the text feature vector, and the hierarchical features are weighted and summed to generate a weighted combined feature. The aggregated features are nonlinearly transformed through an independent residual branch to generate residual supplementary features; The weighted combined features are added to the residual supplementary features, and the result is normalized to generate the multimodal fusion features.

[0012] Preferably, the aggregated features are input into a dynamic weight allocation network, which includes a feedforward layer and a normalization function, and outputs a set of dynamically learned convex combination parameters, including: The aggregated features are input into the feedforward layer of the dynamic weight allocation network. The feedforward layer contains at least one fully connected layer and a ReLU activation function. The aggregated features are then subjected to a nonlinear transformation, and an initial weight vector is output. The initial weight vector is input to a normalization function, which is the Softmax function, to normalize the initial weight vector and ensure that the sum of all its elements is 1, thereby generating a set of convex combination parameters, which include a first weight, a second weight, and a third weight.

[0013] Preferably, the step of inputting the multimodal fusion features into the report generation model to automatically generate a structured analysis report containing data insight conclusions includes: The multimodal fusion features are used as conditions to input into the report generation model based on the Transformer architecture. The report generation model generates a natural language text description containing the data analysis results in an autoregressive manner. The generated natural language text description is integrated with key data metrics extracted from the original structured data to form a structured analysis report containing a title, data summary, key findings, and conclusions / recommendations. The structured analysis report is presented in a visual document format or API interface data format.

[0014] Secondly, this application discloses a BI data analysis device based on an AI model, which adopts the following technical solution, including: The feature extraction module is used to receive analysis requests input by users. The analysis requests include structured data and natural language text. The structured data is encoded using a data feature encoder to obtain a data feature vector. The natural language text is encoded using a text feature encoder to obtain a text feature vector. The feature alignment module is used to extract data sub-features and text sub-features at different levels from multiple network layers of the data feature encoder and the text feature encoder, respectively, based on the data feature vector and the text feature vector. Through a cross-modal attention mechanism, the data sub-features and text sub-features at the same level or adjacent levels are interactively calculated to generate aligned hierarchical features. The feature fusion module is used to aggregate the data feature vector, the text feature vector, and the hierarchical features. Through a dynamic weight allocation network, the contribution weights of the data feature vector, the text feature vector, and the hierarchical features in the fused features are dynamically allocated. Based on the contribution weights, weighted fusion is performed to generate multimodal fused features that can simultaneously reflect data patterns and text semantics. The report generation module is used to input the multimodal fusion features into the report generation model and automatically generate a structured analysis report containing data insight conclusions.

[0015] Thirdly, this application also provides a control device, the device comprising: It includes a memory and a processor, wherein the memory stores computer programs that can be loaded by the processor and executed, such as the AI ​​model-based BI data analysis method described above.

[0016] Fourthly, this application also provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above regarding the AI ​​model-based BI data analysis method.

[0017] In summary, this application receives user analysis requests containing structured data and natural language text; then, a dedicated encoder extracts data feature vectors and text feature vectors respectively; and through a hierarchical feature alignment mechanism, multi-granular sub-features are extracted from different network layers of the encoder. Cross-modal attention is used for hierarchical interactive computation to generate deeply fused hierarchical features. Subsequently, a dynamic weight allocation network adaptively assigns contribution weights to data features, text features, and hierarchical features based on the current query context. Convex combination and residual supplementation generate unified multimodal fusion features. Finally, this feature is input into a Transformer-based report generation model, automatically outputting a structured analysis report containing data insights. Thus, through multi-level semantic alignment and dynamic adaptive fusion, the semantic understanding bias and rigidity of fusion strategies in BI systems are reduced, achieving accurate report generation from multimodal input and significantly improving the accuracy of data analysis. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a BI data analysis method based on an AI model.

[0019] Figure 2 This is a structural block diagram of a BI data analysis device based on an AI model. Detailed Implementation

[0020] The following combination Figure 1 - Figure 2 This application will be described in further detail.

[0021] As enterprise data dimensions become more diversified and analytical needs become more complex, traditional business intelligence systems are struggling to meet the technical challenges of deeply integrating data insights with natural language interaction. While existing solutions attempt to introduce artificial intelligence, their core bottleneck lies in their failure to achieve precise alignment and adaptive fusion of data modalities and text modalities at the deep semantic level. This results in discrepancies between analytical results and user intent, and insufficient responsiveness.

[0022] In view of this, this application proposes a BI data analysis scheme based on multimodal feature alignment and fusion. The core of this application is to solve the cross-modal semantic problem by constructing a hierarchical feature alignment network and a dynamic feature fusion mechanism, so as to realize accurate analysis from raw data and natural language instructions.

[0023] Reference Figure 1 The embodiments of this application include at least steps S10 to S40.

[0024] S10: Receive the analysis request input by the user. The analysis request includes structured data and natural language text. Use a data feature encoder to encode the structured data to obtain a data feature vector. Use a text feature encoder to encode the natural language text to obtain a text feature vector. S20, based on data feature vectors and text feature vectors, extracts data sub-features and text sub-features at different levels from multiple network layers of the data feature encoder and text feature encoder, respectively. Through a cross-modal attention mechanism, it performs interactive calculations on data sub-features and text sub-features at the same level or adjacent levels to generate aligned hierarchical features. S30 aggregates data feature vectors, text feature vectors, and hierarchical features. Through a dynamic weight allocation network, it dynamically allocates the contribution weights of data feature vectors, text feature vectors, and hierarchical features in the fused features. Based on the contribution weights, it performs weighted fusion to generate multimodal fused features that can simultaneously reflect data patterns and text semantics. S40 inputs multimodal fusion features into the report generation model to automatically generate a structured analysis report containing data insight conclusions.

[0025] Specifically, the system simultaneously receives structured data and natural language text, converting them into data feature vectors and text feature vectors respectively through a dedicated encoder. Then, it extracts sub-features at different levels from multiple network layers of the encoder, performing hierarchical interactive computation through a cross-modal attention mechanism to generate deeply fused hierarchical features, effectively solving the semantic gap problem in traditional methods. Subsequently, the original features and hierarchical features are aggregated, and the contribution weights of each feature are adaptively calculated through a dynamic weight allocation network. Weighted fusion generates a multimodal fusion feature that accurately reflects both data patterns and text semantics. Finally, this feature is input into a report generation model, automatically outputting a structured analysis report. Thus, through the synergistic effect of hierarchical alignment and dynamic fusion, the system reduces semantic understanding biases and rigid fusion strategies in BI systems, achieving accurate report generation from multimodal input and significantly improving the accuracy of data analysis.

[0026] In some embodiments, step S10 specifically includes the following steps: preprocessing the received structured data to convert it into a structured data tensor that the model can process; simultaneously, segmenting and encoding the natural language text to convert it into a text tag sequence; inputting the structured data tensor into a data feature encoder, which is a Transformer-based sequence model or graph neural network that encodes the input data through its multi-layer network, and finally outputs a global data feature vector from the last layer; inputting the text tag sequence into a text feature encoder, which is a pre-trained language model that encodes the input sequence through its multi-layer network, and finally outputs a global text feature vector from the last layer; during the forward computation of the data feature encoder and the text feature encoder, simultaneously extracting intermediate representations from the i-th network layer of the data feature encoder as data sub-features, and extracting intermediate representations from the i-th network layer of the text feature encoder as text sub-features; where i is an integer greater than or equal to 1 and less than the total number of network layers.

[0027] Specifically, multimodal feature extraction is achieved through a parallel processing flow: first, structured data is converted into normalized data tensors, and natural language text is segmented and encoded to form a labeled sequence; then, deep feature extraction is performed through a data encoder based on Transformer or graph neural network and a pre-trained language text encoder, respectively. This not only obtains global data feature vectors and text feature vectors from the final output layer, but also extracts the feature representations of each intermediate layer of the network as data sub-features and text sub-features during the encoding process, establishing a multi-level feature foundation for subsequent hierarchical cross-modal alignment, and realizing feature preprocessing from raw data to the synchronous generation of global and local features.

[0028] In some embodiments, step S20 specifically includes the following steps: For the i-th network layer of the encoder, the extracted data sub-features and text sub-features are mapped to a shared d-dimensional semantic space through independent linear transformation layers to obtain mapped data sub-features and mapped text sub-features of a unified dimension; based on the mapped data sub-features and mapped text sub-features, bidirectional cross-modal attention calculation is performed, using the mapped text sub-features as the query and the mapped data sub-features as the key and value, to calculate the attention from text to data and generate text-guided data augmentation features; using the mapped data sub-features as the query and the mapped text sub-features as the key and value, to calculate the attention from data to text and generate data-guided text augmentation features; the data augmentation features and text augmentation features are combined to generate the aligned hierarchical features corresponding to the i-th network layer.

[0029] Furthermore, for the i-th network layer, the data augmentation features and text augmentation features are concatenated to form the primary fusion features of the corresponding network layer; all primary fusion features from all network layers are concatenated to form a comprehensive multi-level fusion representation; the multi-level fusion representation is input into a multilayer perceptron (MLP) for nonlinear transformation and dimensionality reduction, and semantically guided multi-level embedding features are output to obtain hierarchical features.

[0030] First, the data sub-features and text sub-features extracted from each network layer are mapped to a shared semantic space. For the i-th layer, this mapping process can be represented as: ; ; in, and Let W and b be the first hidden states of the image and text encoders in the i-th layer, respectively, and let σ be the learnable projection parameters.

[0031] Then, bidirectional cross-modal attention calculation is performed, and its core calculation formula is as follows: ; This formula calculates attention from text to data and from data to text, generates text-guided data augmentation features and data-guided text augmentation features, and sums them to obtain the fusion output of the i-th layer. .

[0032] Based on this, primary fusion features of each layer are formed through splicing operations. Then, the primary features of all network layers are integrated to construct a multi-level fusion representation. Finally, a multi-layer perceptron is used for nonlinear transformation and dimensionality reduction to output semantically guided embedding features containing multi-level alignment information as hierarchical features. The formula is shown below: ; This establishes a cross-modal alignment system from local details to global semantics, providing a structured feature foundation for subsequent dynamic fusion.

[0033] Specifically, the data sub-features and text sub-features extracted from each network layer are first mapped to a shared semantic space. Then, bidirectional cross-modal attention computation is performed to generate text-guided data augmentation features and data-guided text augmentation features, respectively. On this basis, primary fusion features of each layer are formed through concatenation operations. Then, the primary features of all network layers are integrated to construct a multi-level fusion representation. Finally, a multi-layer perceptron is used for nonlinear transformation and dimensionality reduction to output semantically guided embedding features containing multi-level alignment information as hierarchical features. This establishes a cross-modal alignment system from local details to global semantics, providing a structured feature foundation for subsequent dynamic fusion.

[0034] In some embodiments, step S30 specifically includes the following steps: passing the data feature vector, text feature vector, and hierarchical features through independent linear projection layers, and concatenating the projected features to form aggregated features; inputting the aggregated features into a dynamic weight allocation network, which includes a feedforward layer and a normalization function, and outputs a set of dynamically learned convex combination parameters, including a first weight corresponding to the data feature vector, a second weight corresponding to the text feature vector, and a third weight corresponding to the hierarchical features, and satisfying that the sum of the first weight, the second weight, and the third weight is 1; performing a weighted summation on the projected data feature vector, text feature vector, and hierarchical features based on the first weight, the second weight, and the third weight to generate a weighted combination feature; performing a nonlinear transformation on the aggregated features through an independent residual branch to generate residual supplementary features; adding the weighted combination feature and the residual supplementary feature, and normalizing the result to generate a multimodal fusion feature.

[0035] Furthermore, the aggregated features are input into the feedforward layer of the dynamic weight allocation network. The feedforward layer contains at least one fully connected layer and a ReLU activation function to perform a nonlinear transformation on the aggregated features and output an initial weight vector. The initial weight vector is then input into a normalization function, which is the Softmax function, to normalize the initial weight vector and ensure that the sum of all its elements is 1, generating a set of convex combination parameters, including a first weight, a second weight, and a third weight.

[0036] In specific business scenarios such as sales data analysis, linear projection is used to concatenate data feature vectors, text feature vectors, and hierarchical features into aggregate features. The dynamic weight allocation network generates a set of dynamically learned convex combination parameters α, β, γ based on the current analysis intent through a feedforward layer and Softmax normalization, and satisfies α+β+γ=1.

[0037] Final multimodal fusion features Generate using the following formula: ; in, and These are the projected text and data feature vectors, respectively. As a hierarchical feature, This formula generates supplementary features for the residual branches. By combining weighted summation with residual supplementation, it generates multimodal fusion features that can accurately quantify data patterns and thoroughly understand text semantics, ultimately supporting the generation of intelligent analysis reports.

[0038] In specific business scenarios such as sales data analysis, linear projection is used to concatenate data feature vectors (reflecting sales trends), text feature vectors (user-submitted "analysis of the reasons for the decline in sales of a certain product in East China"), and hierarchical features (deep semantic associations through cross-modal alignment) into aggregate features. The dynamic weight allocation network generates adaptive weights based on the current analysis intent through a feedforward layer and Softmax normalization (such as emphasizing data features to locate the period of decline, strengthening text features to associate with the semantics of user comments, and using hierarchical features to capture the implicit association between regions and products). By combining weighted summation and residual supplementation, multimodal fusion features that can accurately quantify data patterns and thoroughly understand text semantics are generated, ultimately producing an intelligent analysis report.

[0039] In some embodiments, step S40 specifically includes the following steps: inputting multimodal fusion features as conditions into a report generation model based on the Transformer architecture; generating a natural language text description containing data analysis results in an autoregressive manner through the report generation model; integrating the generated natural language text description with key data indicators extracted from the original structured data to assemble a structured analysis report containing a title, data summary, key findings, and conclusions and recommendations; the structured analysis report is presented in a visual document format or API interface data format.

[0040] Specifically, the report generation model based on Transformer uses multimodal fusion features as conditional input. Through autoregressive generation, it transforms data insights into natural language descriptions and combines them with key indicators extracted from the raw data to automatically assemble standardized analysis reports containing a complete logical structure (title, data summary, key findings, and conclusions / recommendations). Finally, it outputs the reports in the form of visual documents or API interfaces, realizing the transformation from complex data features into directly usable analysis reports, significantly improving the usability and visibility of analysis results.

[0041] The implementation principle of a BI data analysis method based on an AI model in this application is as follows: It receives analysis requests from users containing structured data and natural language text; then, a dedicated encoder extracts data feature vectors and text feature vectors respectively; and through a hierarchical feature alignment mechanism, it extracts multi-granularity sub-features from different network layers of the encoder. Cross-modal attention is used for hierarchical interactive computation to generate deeply fused hierarchical features. Next, a dynamic weight allocation network is employed to adaptively allocate the contribution weights of data features, text features, and hierarchical features according to the current query context. A unified multimodal fusion feature is generated through convex combination and residual supplementation. Finally, this feature is input into a Transformer-based report generation model, which automatically outputs a structured analysis report containing data insight conclusions. Thus, through multi-level semantic alignment and dynamic adaptive fusion, it reduces semantic understanding bias and rigidity in BI systems, achieving accurate report generation from multimodal input and significantly improving the accuracy of data analysis.

[0042] Figure 1 This is a flowchart illustrating a BI data analysis method based on an AI model in one embodiment. It should be understood that, although... Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows; unless explicitly stated herein, there is no strict order requirement for the execution of these steps, and they can be executed in other orders; and Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0043] Based on the same technical concept, referring to Figure 2 This application also provides a BI data analysis device based on an AI model, which adopts the following technical solution: The device includes: The feature extraction module is used to receive analysis requests input by users. The analysis requests include structured data and natural language text. The structured data is encoded using a data feature encoder to obtain a data feature vector, and the natural language text is encoded using a text feature encoder to obtain a text feature vector. The feature alignment module is used to extract data sub-features and text sub-features at different levels from multiple network layers of the data feature encoder and text feature encoder, respectively, based on data feature vectors and text feature vectors. Through a cross-modal attention mechanism, it performs interactive calculations on data sub-features and text sub-features at the same level or adjacent levels to generate aligned hierarchical features. The feature fusion module is used to aggregate data feature vectors, text feature vectors, and hierarchical features. Through a dynamic weight allocation network, the contribution weights of data feature vectors, text feature vectors, and hierarchical features in the fused features are dynamically allocated. Based on the contribution weights, weighted fusion is performed to generate multimodal fused features that can simultaneously reflect data patterns and text semantics. The report generation module is used to input multimodal fusion features into the report generation model and automatically generate a structured analysis report containing data insight conclusions.

[0044] In some embodiments, the feature extraction module is specifically used to preprocess the received structured data and convert it into a structured data tensor that the model can process; at the same time, it performs word segmentation and encoding on the natural language text and converts it into a text tag sequence. The structured data tensor is input into the data feature encoder, which is a sequence model or graph neural network based on Transformer. The input data is encoded through its multi-layer network, and the final layer of the network outputs the global data feature vector. The text tag sequence is input into the text feature encoder, which is a pre-trained language model that encodes the input sequence through its multi-layer network. Finally, the last layer of the network outputs the global text feature vector. During the forward computation of the data feature encoder and the text feature encoder, intermediate representations are extracted synchronously from the i-th network layer of the data feature encoder as data sub-features, and intermediate representations are extracted from the i-th network layer of the text feature encoder as text sub-features; where i is an integer greater than or equal to 1 and less than the total number of network layers.

[0045] In some embodiments, the feature alignment module is specifically used to map the extracted data sub-features and text sub-features from the i-th network layer of the encoder to a shared d-dimensional semantic space through independent linear transformation layers, so as to obtain mapped data sub-features and mapped text sub-features of a unified dimension. Based on the mapped data sub-features and mapped text sub-features, bidirectional cross-modal attention computation is performed. Using the mapped text sub-features as the query and the mapped data sub-features as the key and value, text-to-data attention is calculated to generate text-guided data-enhanced features; using the mapped data sub-features as the query and the mapped text sub-features as the key and value, data-to-text attention is calculated to generate data-guided text-enhanced features. The data augmentation features are combined with the text augmentation features to generate the aligned hierarchical features corresponding to the i-th network layer.

[0046] In some embodiments, the feature alignment module is specifically used to concatenate the data augmentation features and the text augmentation features for the i-th network layer to form the primary fusion feature of the corresponding network layer; All primary fusion features from all network layers are spliced ​​together to form a comprehensive multi-level fusion representation; The multi-level fusion representation is input into a multilayer perceptron (MLP) for nonlinear transformation and dimensionality reduction, and the semantically guided multi-level embedding features are output to obtain hierarchical features.

[0047] In some embodiments, the feature fusion module is specifically used to pass the data feature vector, text feature vector and hierarchical features through independent linear projection layers, and then concatenate the projected features to form aggregated features; The aggregated features are input into the dynamic weight allocation network, which includes a feedforward layer and a normalization function. It outputs a set of dynamically learned convex combination parameters, including a first weight corresponding to the data feature vector, a second weight corresponding to the text feature vector, and a third weight corresponding to the hierarchical features, and satisfies that the sum of the first weight, the second weight, and the third weight is 1. Based on the first weight, the second weight, and the third weight, the projected data feature vector, text feature vector, and hierarchical features are weighted and summed to generate a weighted combined feature. The aggregated features are nonlinearly transformed through an independent residual branch to generate residual supplementary features. The weighted combined features are added to the residual supplementary features, and the result is normalized to generate multimodal fusion features.

[0048] In some embodiments, the feature fusion module is specifically used to input aggregated features into the feedforward layer of the dynamic weight allocation network. The feedforward layer includes at least a fully connected layer and a ReLU activation function to perform a nonlinear transformation on the aggregated features and output an initial weight vector. The initial weight vector is input into the normalization function, which is the Softmax function, to normalize the initial weight vector and ensure that the sum of all its elements is 1, generating a set of convex combination parameters, which include the first weight, the second weight, and the third weight.

[0049] In some embodiments, the report generation module is specifically used to input multimodal fusion features as conditions into the report generation module based on the Transformer architecture, and generate a natural language text description containing data analysis results in an autoregressive manner through the report generation model; The generated natural language text description is integrated with key data metrics extracted from the original structured data to form a structured analysis report containing a title, data summary, key findings, and conclusions / recommendations. The structured analysis report is presented in a visual document format or as API interface data.

[0050] This application also discloses a control device.

[0051] Specifically, the control device includes a memory and a processor, with the memory storing computer programs that can be loaded and executed by the processor to perform the aforementioned AI model-based BI data analysis method.

[0052] This application also discloses a computer-readable storage medium.

[0053] Specifically, the computer-readable storage medium stores a computer program that can be loaded by a processor and executed, such as the AI ​​model-based BI data analysis method described above. The computer-readable storage medium includes, for example, various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0054] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A BI data analysis method based on an AI model, characterized in that, include: The system receives an analysis request input by a user, the analysis request including structured data and natural language text. It uses a data feature encoder to encode the structured data to obtain a data feature vector, and uses a text feature encoder to encode the natural language text to obtain a text feature vector. Based on the data feature vector and the text feature vector, data sub-features and text sub-features at different levels are extracted from multiple network layers of the data feature encoder and the text feature encoder, respectively. Through a cross-modal attention mechanism, the data sub-features and text sub-features at the same level or adjacent levels are interactively calculated to generate aligned hierarchical features. The data feature vector, the text feature vector, and the hierarchical features are aggregated, and the contribution weights of the data feature vector, the text feature vector, and the hierarchical features in the fusion feature are dynamically allocated through a dynamic weight allocation network. Based on the contribution weights, weighted fusion is performed to generate a multimodal fusion feature that can simultaneously reflect data patterns and text semantics. The multimodal fusion features are input into the report generation model to automatically generate a structured analysis report containing data insights.

2. The BI data analysis method based on an AI model according to claim 1, characterized in that, The analysis request received from the user includes structured data and natural language text. A data feature encoder is used to encode the structured data to obtain a data feature vector, and a text feature encoder is used to encode the natural language text to obtain a text feature vector, including: The received structured data is preprocessed and converted into a structured data tensor that the model can process; simultaneously, the natural language text is segmented and encoded, and converted into a text tag sequence. The structured data tensor is input to the data feature encoder, which is a sequence model or graph neural network based on Transformer. The input data is encoded through its multi-layer network, and the last layer of the network outputs the global data feature vector. The text tag sequence is input into the text feature encoder, which is a pre-trained language model that encodes the input sequence through its multi-layer network, and finally outputs the global text feature vector by the last layer of the network. During the forward computation of the data feature encoder and the text feature encoder, intermediate representations are extracted synchronously from the i-th network layer of the data feature encoder as the data sub-features, and intermediate representations are extracted from the i-th network layer of the text feature encoder as the text sub-features; where i is an integer greater than or equal to 1 and less than the total number of network layers.

3. The BI data analysis method based on an AI model according to claim 2, characterized in that, Based on the data feature vector and the text feature vector, different levels of data sub-features and text sub-features are extracted from multiple network layers of the data feature encoder and the text feature encoder, respectively. Through a cross-modal attention mechanism, the data sub-features and text sub-features at the same or adjacent levels are interactively calculated to generate aligned hierarchical features, including: For the i-th network layer of the encoder, the extracted data sub-features and text sub-features are respectively mapped to a shared d-dimensional semantic space through independent linear transformation layers to obtain mapped data sub-features and mapped text sub-features of a unified dimension. Based on the mapped data sub-features and mapped text sub-features, bidirectional cross-modal attention calculation is performed. Using the mapped text sub-features as the query and the mapped data sub-features as the key and value, text-to-data attention is calculated to generate text-guided data augmentation features. Using the mapped data sub-features as the query and the mapped text sub-features as the key and value, data-to-text attention is calculated to generate data-guided text augmentation features. The data augmentation features are combined with the text augmentation features to generate the aligned hierarchical features corresponding to the i-th network layer.

4. The BI data analysis method based on an AI model according to claim 3, characterized in that, The step of combining the data augmentation features with the text augmentation features to generate the aligned hierarchical features corresponding to the i-th network layer includes: For the i-th network layer, the data augmentation feature and the text augmentation feature are concatenated to form the primary fusion feature of the corresponding network layer; All the primary fusion features from all network layers are spliced ​​together to form a comprehensive multi-level fusion representation; The multi-level fusion representation is input into a multilayer perceptron (MLP) for nonlinear transformation and dimensionality reduction, and semantically guided multi-level embedding features are output to obtain the hierarchical features.

5. The BI data analysis method based on an AI model according to claim 1, characterized in that, The process involves aggregating the data feature vector, the text feature vector, and the hierarchical features; dynamically allocating the contribution weights of the data feature vector, the text feature vector, and the hierarchical features in the fused feature using a dynamic weight allocation network; and performing weighted fusion based on the contribution weights to generate a multimodal fused feature that simultaneously reflects data patterns and text semantics. The data feature vector, the text feature vector, and the hierarchical features are each passed through an independent linear projection layer, and the projected features are then concatenated to form aggregated features. The aggregated features are input into a dynamic weight allocation network, which includes a feedforward layer and a normalization function, and outputs a set of dynamically learned convex combination parameters, including a first weight corresponding to the data feature vector, a second weight corresponding to the text feature vector, and a third weight corresponding to the hierarchical features, and the sum of the first weight, the second weight, and the third weight is 1. Based on the first weight, the second weight, and the third weight, the projected data feature vector, the text feature vector, and the hierarchical features are weighted and summed to generate a weighted combined feature. The aggregated features are nonlinearly transformed through an independent residual branch to generate residual supplementary features; The weighted combined features are added to the residual supplementary features, and the result is normalized to generate the multimodal fusion features.

6. The BI data analysis method based on an AI model according to claim 5, characterized in that, The aggregated features are input into a dynamic weight allocation network, which includes a feedforward layer and a normalization function, and outputs a set of dynamically learned convex combination parameters, including: The aggregated features are input into the feedforward layer of the dynamic weight allocation network. The feedforward layer contains at least one fully connected layer and a ReLU activation function. The aggregated features are then subjected to a nonlinear transformation, and an initial weight vector is output. The initial weight vector is input to a normalization function, which is the Softmax function, to normalize the initial weight vector and ensure that the sum of all its elements is 1, thereby generating a set of convex combination parameters, which include a first weight, a second weight, and a third weight.

7. The BI data analysis method based on an AI model according to claim 1, characterized in that, The step of inputting the multimodal fusion features into the report generation model to automatically generate a structured analysis report containing data insight conclusions includes: The multimodal fusion features are used as conditions to input into the report generation model based on the Transformer architecture. The report generation model generates a natural language text description containing the data analysis results in an autoregressive manner. The generated natural language text description is integrated with key data metrics extracted from the original structured data to form a structured analysis report containing a title, data summary, key findings, and conclusions / recommendations. The structured analysis report is presented in a visual document format or API interface data format.

8. A BI data analysis device based on an AI model, characterized in that, The device includes: The feature extraction module is used to receive analysis requests input by users. The analysis requests include structured data and natural language text. The structured data is encoded using a data feature encoder to obtain a data feature vector. The natural language text is encoded using a text feature encoder to obtain a text feature vector. The feature alignment module is used to extract data sub-features and text sub-features at different levels from multiple network layers of the data feature encoder and the text feature encoder, respectively, based on the data feature vector and the text feature vector. Through a cross-modal attention mechanism, the data sub-features and text sub-features at the same level or adjacent levels are interactively calculated to generate aligned hierarchical features. The feature fusion module is used to aggregate the data feature vector, the text feature vector, and the hierarchical features. Through a dynamic weight allocation network, the contribution weights of the data feature vector, the text feature vector, and the hierarchical features in the fused features are dynamically allocated. Based on the contribution weights, weighted fusion is performed to generate multimodal fused features that can simultaneously reflect data patterns and text semantics. The report generation module is used to input the multimodal fusion features into the report generation model and automatically generate a structured analysis report containing data insight conclusions.

9. A control device, characterized in that, The device includes: A memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 7.

Images