AI service information acquisition method

Abstract

The invention relates to the technical field of artificial intelligence, and discloses an AI service information acquisition method. The method comprises the following steps: S1, receiving multi-modal input data from different sources; s2, preprocessing input data in the comprehensive data structure; s3, constructing a dynamic graph neural network based on the embedded representation; s4, updating the node state by using an adaptive solver; s5, carrying out four-dimensional tensor online decomposition on the preprocessed multi-mode embedding; s6, dynamically updating the core tensor through newly added data; s7, cross-modal alignment is realized; and S8, providing an online service interface. And by introducing a mechanism of dynamically updating the core tensor, the capability of the system to adapt to different environments is remarkably improved. Compared with a model with fixed parameters in the prior art, the scheme can be continuously adjusted according to user feedback and real-time data, so that the system is more flexible and effectively responds to changes of user requirements. The self-adaptive characteristic ensures that the user experience is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and specifically to a method for obtaining AI service information. Background Art

[0002] Existing technologies typically rely on static feature alignment to process multimodal data. In practice, this approach often results in poor matching results and fails to accurately reflect the inherent relationships between different modal information. The limitations of static features make it difficult for systems to flexibly respond to dynamically changing user needs, impacting the user experience.

[0003] Many legacy systems struggle with parameter updates, often relying on fixed parameters. This fixed nature makes the systems less agile when dealing with real-time data. User needs change rapidly, and these technologies fail to adapt in a timely manner, resulting in delayed information response and an inability to provide effective services.

[0004] Furthermore, existing online service interfaces often only support a single input modality. This singleness limits user interaction methods, forcing users to adhere to a fixed input format when searching for information. This clearly fails to meet the increasingly diverse needs of modern users. The lack of comprehensive processing of multimodal information compromises both service accuracy and timeliness.

[0005] Furthermore, current technologies often rely on simple feature concatenation when processing multimodal information fusion. This approach fails to fully consider the relative importance of each modality, resulting in a lack of clear relevance and context in the final feedback information. This significantly reduces the efficiency and effectiveness of user information acquisition. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention provides an AI service information acquisition method, which solves the problems of insufficient dynamic adaptability, low cross-modal matching accuracy, and poor real-time feedback efficiency in multimodal information processing in the existing technology.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: A method for obtaining AI service information, comprising: S1, receives multimodal input data from different sources, including text, images, and time series data, and aggregates them into a comprehensive data structure; S2. Preprocess the input data in the comprehensive data structure to generate embedding representations of the respective modalities; S3. Build a dynamic graph neural network based on embedded representation and control the evolution of node states through differential equations to achieve real-time capture of data streams. S4. Use the adaptive solver to update the node state to reflect the changes in the time dimension, and feed the updated results back to the dynamic graph neural network; S5. Perform online four-dimensional tensor decomposition on the preprocessed multimodal embedding to process the interaction between dimensions and generate core tensors and factor matrices. S6. The core tensor is dynamically updated with the newly added data. The updated core tensor will be used for subsequent modality alignment. S7. Implement cross-modal alignment by computing hyperbolic space projections to enhance the matching between different modalities and using the updated core tensor in this process; S8. Provide an online service interface to receive user queries and feedback relevant information in real time. When users query, real-time feedback is provided through processed multimodal embeddings and updated core tensors.

[0008] Preferably, the preprocessing step includes encoding the text modality using a dynamically quantized BERT model, processing the image modality using a pruned VisionTransformer, and encoding the time series data using causal convolution.

[0009] Preferably, the differential equation used in the constructed dynamic graph neural network is: ; in, Representation node The time derivative of the hidden state of represents the activation function, represents the weight matrix, Represents an aggregate function, Represents neighbor nodes In the Hidden state vector in layer graph neural network, Representation node The set of neighbor nodes of represents the hyperparameter, Representation node The dynamic state of Representation node The time derivative of the dynamic state of .

[0010] Preferably, the adaptive solver adopts an iterative algorithm with adaptive step size to update the node status.

[0011] Preferably, the four-dimensional tensor online decomposition calculation is achieved by the following objective function: ; in, Represents the input four-dimensional tensor data, , , , , Indicates that these are the variables in the optimization process, represents the Frobenius norm, represents the regularization parameter, Representing a tensor At the time point slices, Representing a tensor At the time point slices, Indicates the accumulation from time point 2 to The error term represents the difference between calculation time points.

[0012] Preferably, the process of dynamically updating the core tensor comprises the following steps: Receive new time slice data and compare it with the current core tensor to determine the changes in the data; According to the preset update rules, the weighted average method is applied to fuse the new data with the current core tensor to generate an updated core tensor.

[0013] Preferably, in the weighted averaging method, the weight coefficient is determined by the importance of the new data and the stability of the previous core tensor, so as to ensure that the updated core tensor reflects the real changes of the data and retains historical information.

[0014] Preferably, the cross-modal alignment is achieved through hyperbolic space, and its loss function is: ; in, represents the cross-modal loss function, For all matching pairs Perform the summation, represents the inverse hyperbolic cosine function, Represents the source sample and target samples The inner product of represents the eigenvector of the source modality sample, represents the feature vector of the target modality sample, Represents the source sample feature vector The square norm of Represents the target sample feature vector The square norm of represents the hyperparameter, represents the Kullback-Leibler divergence, represents the predicted probability distribution of the model, represents the target distribution, usually refers to the true label distribution of the target modality.

[0015] Preferably, the online service interface is capable of receiving user queries and returning relevant information through real-time feedback, which forms a number of recommended services.

[0016] Preferably, the multimodal input data includes text input by the user, image data from the sensor, and timing signals acquired in real time, and the signals ensure the accuracy and timeliness of information acquisition.

[0017] The present invention provides a method for obtaining AI service information. It has the following beneficial effects: 1. This invention significantly improves the system's ability to adapt to diverse environments by introducing a mechanism for dynamically updating core tensors. Compared to existing models with fixed parameters, this solution can continuously adjust based on user feedback and real-time data, making the system more flexible and effectively responding to changing user needs. This adaptive nature ensures a significantly improved user experience.

[0018] 2. This invention optimizes multimodal information matching by combining hyperbolic space projection technology with cross-modal alignment. Compared to existing approaches that rely solely on static feature alignment, this invention effectively addresses the lack of similarity in multimodal information processing. This enables the features of each modality to more accurately reflect the real content, achieving higher-quality service information acquisition.

[0019] 3. By building a comprehensive online service interface, this invention achieves an efficient information feedback process. Compared to traditional interfaces that can only handle single-modal queries, this interface can receive multimodal input in real time and provide accurate information feedback. This comprehensive integration capability effectively enhances the user's interactive experience and meets their more complex information needs.

[0020] 4. This invention utilizes a weighted fusion approach to process information from different modalities, thereby optimizing the effectiveness of feature splicing. Compared to the simple, mechanical feature splicing used in existing technologies, this innovative approach fully considers the contribution of each modality, making the final feedback information more relevant. This technology improves the overall accuracy and applicability of information, making multimodal information transmission more fluid and natural. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0022] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0023] Please see the attached Figure 1 , an embodiment of the present invention provides an AI service information acquisition method, including: S1, receives multimodal input data from different sources, including text, images, and time series data, and aggregates them into a comprehensive data structure; Specifically, in step S1 of this embodiment, a comprehensive receiving model for multimodal input data is established. The model is used to receive multimodal data from different data sources, which may include text, images, and time series data. The input from each data source is uniformly transmitted to the model through a defined data preprocessing interface.

[0024] Specifically, text data is input into the system by users and preprocessed using Natural Language Processing (NLP) technology to extract semantic information. Image data is collected by cameras or sensors and initially analyzed using image recognition algorithms. Time series data comes from various sensors and can reflect changes in the situation.

[0025] During the ingestion process, the model first converts each data type into a structured format suitable for subsequent operations. Text data undergoes word segmentation, breaking sentences down into words or phrases. Image data is converted into feature vectors using convolutional neural networks (CNNs), which extract key information from the image. Time series data undergoes data normalization to ensure consistency within numerical ranges.

[0026] The multimodal data can be represented as the following comprehensive data structure: ; in, Represents text data, represents image data, Represents time series data. Each part of the data is in the overall structure The same identifier is used to facilitate subsequent data processing.

[0027] Next, by using the data storage module, the received modal data is stored in the same data warehouse to achieve centralized data management. At the same time, the connection relationship and data flow of each modal data storage module are as follows: The text modality input module is connected to the text data preprocessing module; The image modality input module is connected to the image data processing module; The time series data input module is connected to the time series data normalization module.

[0028] The stored and processed data will form a comprehensive data structure to facilitate the operation of subsequent steps. At this stage, the data is integrated through a unified interface to ensure the integrity and stability of the system.

[0029] By centrally integrating multimodal data sources, the information acquisition system can quickly respond and process data in a dynamically changing environment, improving data consistency and accuracy. Furthermore, the model can adapt to inputs of different data types to effectively capture diverse information in complex scenarios.

[0030] S2. Preprocess the input data in the comprehensive data structure to generate embedding representations of the respective modalities; Specifically, in the present invention, the comprehensive structure of multimodal input data established in step S1 lays the foundation for subsequent data preprocessing. Step S2 involves preprocessing the input data in the comprehensive data structure to generate embedded representations for each modality. This process involves specific algorithms for text, image, and time series data, ensuring that information from different modalities can be effectively captured and encoded.

[0031] In this embodiment, the dynamically quantized BERT (Bidirectional Encoder Representations from Transformers) model is used for encoding for text modalities. This model has deep learning capabilities and can generate vector representations of words based on context. The text data is processed through word segmentation and input into the BERT model. The generated embedding representation is: ; in, For the generated text embedding representation, Represents input text data. The model is able to utilize bidirectional contextual information and significantly improve the accuracy of text embedding.

[0032] When processing image modalities, this embodiment uses the pruned VisionTransformer. Specifically, The image can be divided into fixed-size tiles, and features are extracted for each tile. The model uses a self-attention mechanism, and the output image features can be expressed as: ; in, is the generated image embedding representation, Represents the input image data. Through pruning operation, The computational complexity is reduced, the processing efficiency is improved, and the system can process larger-scale image data.

[0033] For time series data, this embodiment introduces causal convolution to extract important features from the time series. Causal convolution ensures that each time step of the output depends only on the current and previous inputs, thereby protecting the causal relationship of the time series. The mathematical expression of time series data processing is: ; in, To embed the generated time series data, Represents the input time series data, Represents a causal convolution operation. Causal convolution can avoid information leakage and increase the accuracy of the model in time prediction.

[0034] Integrating the embedding representations of each modality, the final result of step S2 is a comprehensive embedding feature matrix: ; in, Represents the merged embedding feature matrix, which contains rich information of text, image and time series data.

[0035] Based on the comprehensive data structure and after the aforementioned preprocessing, the features of each modality can be effectively input into the subsequent graph neural network model for training and inference. At the same time, step S2 ensures the full representation of multimodal data and adapts to the needs of integrating features from different modalities.

[0036] S3. Build a dynamic graph neural network based on embedded representation and control the evolution of node states through differential equations to achieve real-time capture of data streams. Specifically, in step S3 of this embodiment, a dynamic graph neural network (DGN) is constructed based on the multimodal embedding representation generated in step S2. This network controls the evolution of node states through differential equations, thereby capturing data streams in real time. Through this process, the system can effectively process data from different modalities and adapt to rapidly changing environments.

[0037] In this embodiment, the core of constructing DGN is to capture information flow by utilizing the dynamic evolution of node states. Here, the node state vector can be expressed as: ; in, Indicates time moment, node status, Representation node The initial state, that is, at time The state of the moment, Representation node The time derivative of the state indicates the rate of change of the node state over time. From time 0 to The total amount of state changes, reflecting the cumulative changes of the state over time.

[0038] In the design of differential equations, the following equations are introduced in this embodiment to reflect the state changes of nodes in the neural network: ; in, Representation node The time derivative of the hidden state of represents the activation function, represents the weight matrix, Represents an aggregate function, Represents neighbor nodes In the Hidden state vector in layer graph neural network, Representation node The set of neighbor nodes of represents a hyperparameter that controls the influence of the second dynamic change on the total rate of change. Representation node The dynamic state of Representation node The time derivative of the dynamic state of a network, i.e., the speed at which these physical / semantic properties change.

[0039] In this possible implementation, the relationship between nodes is realized through an aggregation function (AGG), which allows the model to collect information from neighboring nodes to update the state of the current node. This process allows the features of various modalities to be aggregated to form a richer node representation.

[0040] In the process of building a DGN, the aforementioned modeling approach allows for the effective integration and utilization of data from various modalities, while the choice of aggregation function makes the network flexible and adaptable. This approach effectively addresses the problem of multimodal data fusion and lays the theoretical foundation for the real-time and accurate nature of information acquisition systems.

[0041] S4. Use the adaptive solver to update the node state to reflect the changes in the time dimension, and feed the updated results back to the dynamic graph neural network; Specifically, in step S4 of this embodiment, an adaptive solver is used to update the node states in the dynamic graph neural network to effectively reflect changes in the time dimension. In this way, the dynamic characteristics of the real-time data stream can be captured and fed back to the network, ensuring that the system can always respond to the latest conditions.

[0042] In this embodiment, an iterative algorithm with adaptive step size is used to improve the efficiency and accuracy of node state updates. Specifically, the node state update is performed based on a differential equation, which is in the form of: ; in, Indicates time When the node status, Indicates time When the node status, represents the time step, which is used for state update in discrete time, Representation node The time derivative of the state.

[0043] In this possible implementation, the adaptive solver can dynamically adjust the In general, when the system changes rapidly, the adaptive step size will be smaller to improve the update accuracy; when the system changes slowly, the step size can be appropriately increased to improve the computational efficiency.

[0044] To better implement adaptive updates, this embodiment introduces the following update strategies: ; in, represents the current time step, represents the maximum allowed time step, represents the adjustment factor, represents some rate of change or derivative, Represents a small constant that prevents division by zero and ensures numerical stability.

[0045] Specifically, this adaptive solver automatically adjusts the step size based on the rate of change of each node's state. By monitoring node state changes in real time, if a dramatic change is detected, the step size is reduced to ensure update reliability; conversely, if the change is stable, the step size is increased to improve computational efficiency. This approach ensures that the system exhibits excellent stability and accuracy under various conditions.

[0046] Furthermore, the results of this update mechanism are fed back into the dynamic graph neural network, ensuring that the newly constructed node states are promptly applied to influence subsequent processing. The node state update formula recalculates all node information in the network by substituting the new state, thus achieving a dynamic flow of information.

[0047] S5. Perform online four-dimensional tensor decomposition on the preprocessed multimodal embedding to process the interaction between dimensions and generate core tensors and factor matrices. Specifically, in step S5 of this embodiment, the node state updated in step S4 is subjected to online four-dimensional tensor decomposition. This process aims to deeply explore the potential relationships and interactive features between multimodal data and integrate information from different modalities based on the core tensor to achieve higher-level data analysis and information extraction.

[0048] In this embodiment, a four-dimensional tensor is first constructed from the node state vector obtained in step S4. Assume that the embedded feature vector corresponding to each mode can be expressed as: ; in, Indicates at a point in time The comprehensive four-dimensional tensor of Represent the embedded features of text, image and time series modalities respectively, and Represents time-related feature information.

[0049] In this process, the decomposition of the four-dimensional tensor can be expressed as: ; in, Represents the input four-dimensional tensor data, , , , , Indicates that these are the variables in the optimization process, represents the Frobenius norm, represents the regularization parameter, Representing a tensor At the time point slices, Representing a tensor At the time point slices, Indicates the accumulation from time point 2 to The error term represents the difference between calculation time points.

[0050] In some embodiments, both core tensor updates and loss optimization are achieved through gradient descent. Specifically, the loss function is calculated through an iterative optimization algorithm, and the factor matrix is ​​updated using backpropagation techniques. The optimized core tensor is able to represent the most important features in the data. This design ensures the effective integration of basic features and complex interactions.

[0051] In this possible implementation, four-dimensional tensor decomposition can not only effectively capture the broad correlation between data, but also reveal the deep connection between different modalities. This process provides an important foundation for the processing of subsequent modules and also facilitates subsequent data fusion and feature extraction. Furthermore, through real-time decomposition and updating, the model can be quickly adjusted as new data arrives, improving the system's flexibility in dynamically changing scenarios. This strategy significantly enhances the system's adaptability to a variety of scenarios, making real-time information acquisition more efficient and reliable.

[0052] S6. The core tensor is dynamically updated with the newly added data. The updated core tensor will be used for subsequent modality alignment. Specifically, in step S6 of this embodiment, intelligent reasoning and predictive analysis are performed based on the features of the four-dimensional tensor decomposed in step S5. This process aims to provide more accurate decision support and information acquisition capabilities by integrating features across different modalities. This step is closely linked to the previous steps, ensuring the effective utilization of multimodal information through in-depth data analysis.

[0053] In this embodiment, the core tensor constructed previously is used With factor matrix , to reason about multimodal data. The details are as follows: ; Here, Indicates at a point in time The predicted output of Represents the tensor multiplication along the nth dimension. Through this formula, the predicted four-dimensional tensor can be reconstructed to extract the important complementary information between the modalities.

[0054] In this process, specifically, in some embodiments, the weighted average method is applied to fuse the features of different modalities. The reason for using weights is that the influence of each modality on the final decision may be different, and a weight vector is set ,in 、 、 are the weights of text, image and time series data respectively. The fusion calculation can be expressed as; ; in, represents a weight vector, represents the weight of the text, represents the weight of the image, Indicates the weight of time data, indicating the importance of time data in the final decision, Represents text features, Represents image features, Represents time characteristics, Represents the final fusion result, which represents the multimodal features after weighted summation fusion In this way, by adjusting the weights, the contribution of different data modalities to the results can be dynamically optimized.

[0055] In one possible implementation, the system inputs the fused features (FF) into a decision module, which can be a classifier or regression model, to implement specific application scenarios. For example, in information retrieval or recommendation systems, further analysis of the features can help the system determine user needs and make corresponding recommendations.

[0056] In addition, the system can also use collaborative filtering technology to further optimize the decision-making process through the user's historical behavior data. Assuming that RR is the interaction matrix between the user and the information item, the user's predicted score for a piece of information can be expressed as: ; in, is the user feature vector, is the information item feature vector, For users Information Item The predicted score of .

[0057] Through continuous iteration and optimization, the system can adjust its decisions in a timely manner to adapt to changing user needs. This strategy makes intelligent services more flexible and personalized, ultimately achieving higher user satisfaction.

[0058] In summary, step S6 ensures the efficiency and adaptability of the intelligent system in dynamic environments through multimodal information extraction and decision support based on four-dimensional tensors. This process transforms the ability to fuse different modalities into practical applications, providing users with accurate and efficient service information.

[0059] S7. Implement cross-modal alignment by computing hyperbolic space projections to enhance the matching between different modalities and using the updated core tensor in this process; Specifically, in step S7 of this embodiment, cross-modal alignment is implemented by computing hyperbolic space projections to enhance the compatibility between different modalities. This process, based on the core tensor constructed in step S6, aims to improve the representation consistency and interrelationships of multimodal data, thereby providing more accurate intelligent service information in practical applications.

[0060] In this embodiment, the core of cross-modal alignment is to construct an effective loss function to quantify the similarity between source and target modality samples. The cross-modal loss function is set as follows: ; in, represents the cross-modal loss function, For all matching pairs Perform the summation, Represents the inverse hyperbolic cosine function, used to calculate the source sample and target samples The similarity between Represents the source sample and target samples The inner product of represents the eigenvector of the source modality sample, represents the feature vector of the target modality sample, Represents the source sample feature vector The square norm of , which represents the size of the source sample, Represents the target sample feature vector The square norm of , which represents the size of the target sample, represents the hyperparameter, represents the Kullback-Leibler divergence, represents the predicted probability distribution of the model, represents the target distribution, usually refers to the true label distribution of the target modality.

[0061] Specifically, the embedding representation of the source modality sample can be described as: ; The embedding representation of the target modality sample is: ; in, The eigenvector of the source modal sample, The feature vector of the target modality sample, Core Matrix, and are the factor matrices of the source mode and the target mode, respectively. The core tensor here is Contains deep features of multimodal data.

[0062] The advantage of using the inverse hyperbolic cosine function in step S7 is that it can effectively handle highly nonlinear features. This function allows the distance metric between data points in hyperbolic space to differ from that in traditional Euclidean space, better capturing the complex relationship between the source and target modalities, especially when the data is difficult to map in traditional space.

[0063] To further enhance the matching, choose appropriate hyperparameters is crucial. Generally speaking, The value of can be optimized by cross-validation to achieve hyperbolic loss and The optimal balance of divergence loss. This trade-off ensures that the heterogeneity between features of different modalities is properly handled during the alignment process.

[0064] Furthermore, the matching pairs included in the core pair loss calculation can be dynamically generated based on the four-dimensional tensor decomposition results obtained in the previous step S5. This means that as the system's ability to process new data improves, the cross-modal alignment objectives can be continuously adjusted and optimized. For example, if a new sample pair is identified as a suitable matching pair, the system can prioritize its training to improve the accuracy and reliability of the overall model.

[0065] S8. Provide an online service interface to receive user queries and feedback relevant information in real time. When users query, real-time feedback is provided through processed multimodal embeddings and updated core tensors.

[0066] Specifically, in step S8, an online service interface is provided to receive user queries and feedback relevant information in real time. This interface uses the previously processed multimodal embedding and updated core tensors to provide users with efficient and accurate feedback information.

[0067] First, an online service interface is established, which includes a user query input module, an information processing module, and an information feedback module. The user query input module receives query requests from users. These requests can be text, voice, or other multimodal inputs. The input query information is preprocessed and converted into multimodal embedding features.

[0068] The information processing module uses the core tensor constructed in step S5 , combined with processed multimodal embedding , actually processing the user query. This process leverages previously implemented cross-modal alignment technology to ensure accurate alignment of information across modalities. Specifically, the information processing module first searches the core tensor for features relevant to the user query and then performs a weighted aggregation on these features.

[0069] For example, the weighted embedding representations of different modalities are calculated by the following formula: ; in, are the weight coefficients of text, image and time modalities respectively, is the corresponding multimodal embedding feature.

[0070] Finally, the information processing module generates the most relevant feedback for the user query by verifying the similarity of the core tensors. The cross-modal loss function defined in step S7 is used to further optimize the match between the source and target modalities. This process ensures that the feedback provided is highly relevant and accurate.

[0071] Finally, the feedback module provides the generated feedback to the user in an appropriate format. This feedback can be in the form of text, images, or other multimedia to ensure that the user can easily understand the information provided. After receiving user interaction feedback, this module can use this feedback to continuously optimize the content and method of message delivery, thereby improving the user experience.

[0072] In this process, the combination of multimodal embedding and core tensors strengthens the logical connection between user queries and tasks within the information processing module. This enables efficient information flow within the information module to support dynamic query processing. Every user query is processed through this logical chain, ensuring the real-time and accurate feedback.

[0073] In this implementation, the system can support highly concurrent user queries and update core tensors and model parameters in the backend in real time to adapt to changing user needs. This ensures that the system can provide intelligent services to users in a sustainable and effective manner.

[0074] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for obtaining AI service information, characterized in that: include; S1, receives multimodal input data from different sources, including text, images, and time series data, and aggregates them into a comprehensive data structure; S2. Preprocess the input data in the comprehensive data structure to generate embedding representations of the respective modalities; S3. Build a dynamic graph neural network based on embedded representation and control the evolution of node states through differential equations to achieve real-time capture of data streams. S4. Use the adaptive solver to update the node state to reflect the changes in the time dimension, and feed the updated results back to the dynamic graph neural network; S5. Perform online four-dimensional tensor decomposition on the preprocessed multimodal embedding to process the interaction between dimensions and generate core tensors and factor matrices. S6. The core tensor is dynamically updated with the newly added data. The updated core tensor will be used for subsequent modality alignment. S7. Implement cross-modal alignment by computing hyperbolic space projections to enhance the matching between different modalities and using the updated core tensor in this process; S8. Provide an online service interface to receive user queries and feedback relevant information in real time. When users query, real-time feedback is provided through processed multimodal embeddings and updated core tensors.

2. The AI ​​service information acquisition method according to claim 1, characterized in that: The preprocessing steps include encoding the text modality using a dynamically quantized BERT model, processing the image modality using a pruned VisionTransformer, and encoding the time series data using causal convolution.

3. The AI ​​service information acquisition method according to claim 1, characterized in that: The differential equation used in the constructed dynamic graph neural network is: ； in, Representation node The time derivative of the hidden state of represents the activation function, represents the weight matrix, Represents an aggregate function, Represents neighbor nodes In the Hidden state vector in layer graph neural network, Representation node The set of neighbor nodes of represents the hyperparameter, Representation node The dynamic state of Representation node The time derivative of the dynamic state of .

4. The AI ​​service information acquisition method according to claim 1, characterized in that: The adaptive solver uses an iterative algorithm with adaptive step size to update node states.

5. The AI ​​service information acquisition method according to claim 1, characterized in that: The four-dimensional tensor online decomposition calculation is achieved through the following objective function: ； in, Represents the input four-dimensional tensor data, , , , , Indicates that these are the variables in the optimization process, represents the Frobenius norm, represents the regularization parameter, Representing a tensor At the time point slices, Representing a tensor At the time point slices, Indicates the accumulation from time point 2 to The error term represents the difference between calculation time points.

6. The AI ​​service information acquisition method according to claim 1, characterized in that: The process of dynamically updating the core tensor includes the following steps: Receive new time slice data and compare it with the current core tensor to determine the changes in the data; According to the preset update rules, the weighted average method is applied to fuse the new data with the current core tensor to generate an updated core tensor.

7. The AI ​​service information acquisition method according to claim 6, characterized in that: In the weighted averaging method, the weight coefficient is determined by the importance of the new data and the stability of the previous core tensor, so as to ensure that the updated core tensor reflects the real changes of the data and retains historical information.

8. The AI ​​service information acquisition method according to claim 1, characterized in that: The cross-modal alignment is achieved through hyperbolic space, and its loss function is: ； in, represents the cross-modal loss function, For all matching pairs Perform the summation, represents the inverse hyperbolic cosine function, Represents the source sample and target samples The inner product of represents the eigenvector of the source modality sample, represents the feature vector of the target modality sample, Represents the source sample feature vector The square norm of Represents the target sample feature vector The square norm of represents the hyperparameter, represents the Kullback-Leibler divergence, represents the predicted probability distribution of the model, represents the target distribution, usually refers to the true label distribution of the target modality.

9. The AI ​​service information acquisition method according to claim 1, characterized in that: The online service interface is capable of receiving user queries and returning relevant information through real-time feedback, which forms a number of recommended services.

10. The AI ​​service information acquisition method according to claim 1, characterized in that: The multimodal input data includes text input by the user, image data from the sensor, and timing signals acquired in real time, and the signals ensure the accuracy and timeliness of information acquisition.

Images