Network planning intention recognition method, device, equipment, medium and product

Abstract

The invention relates to a network planning intention recognition method and device, equipment, a medium and a product. The method comprises the following steps: acquiring data of at least two modes of a network to be identified; inputting the data of the at least two modalities into a pre-trained intention recognition model, performing joint semantic enhancement on the data of all the modalities, and performing attention fusion on the features of the first type of data in the data of the at least two modalities, the attention fusion result, the joint semantic enhancement result and features of second type data in the data of the at least two modalities are subjected to dynamic balance processing to obtain a target data result, and the data of the at least two modalities are divided based on a preset division rule to obtain first type data and the second type data; and identifying the target data result to obtain a network planning intention identification result corresponding to the to-be-identified network. According to the invention, the accuracy of network planning intention recognition can be improved.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and communication network technology, and in particular to a method, apparatus, device, medium and product for identifying network planning intentions. Background Technology

[0002] With the accelerated development of communication technologies and cloud-network convergence, planning for networks such as the Internet and the Internet of Things is a very common practice, and network planning can improve network performance.

[0003] Network planning typically requires execution based on identified network planning intentions. In related technologies, network planning intention identification usually involves feature analysis of data from multiple network modalities to determine the intention. However, these technologies suffer from low accuracy in identifying network planning intentions based on multimodal data. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, device, medium, and product for identifying network planning intent in response to the above-mentioned technical problems.

[0005] Firstly, this application provides a method for identifying network planning intent, including:

[0006] Acquire data for at least two modalities of the network to be identified;

[0007] At least two modalities of data are input into a pre-trained intent recognition model. Joint semantic enhancement is performed on all modalities of data. Attention fusion is performed on the features of the first type of data in the at least two modalities of data. The attention fusion result, the joint semantic enhancement result, and the features of the second type of data in the at least two modalities of data are dynamically balanced to obtain the target data result. Among them, the at least two modalities of data are divided according to a preset partitioning rule to obtain the first type of data and the second type of data.

[0008] The target data results are identified to obtain the network planning intent identification results corresponding to the network to be identified.

[0009] In one embodiment, the intent recognition model includes a self-evolving semantic module, a quantum fusion module, and a gated fusion module; data from at least two modalities are input into the pre-trained intent recognition model, joint semantic enhancement is performed on the data from all modalities, attention fusion is performed on the features of the first type of data in the data from at least two modalities, and the attention fusion result, the joint semantic enhancement result, and the features of the second type of data in the data from at least two modalities are dynamically balanced to obtain the target data result, including:

[0010] Joint semantic enhancement is performed on data from all modalities through a self-evolving semantic module;

[0011] Attention fusion of features of first-type data from at least two modalities is performed using a quantum fusion module.

[0012] The gating fusion module dynamically balances the attention fusion result, the joint semantic enhancement result, and the features of the second type of data from at least two modalities to obtain the target data result.

[0013] In one embodiment, the intent recognition model further includes an intent decoupling module, which identifies the target data results to obtain network planning intent recognition results corresponding to the network to be identified, including:

[0014] The intent decoupling module segments the target data results into preset dimension regions to obtain network intent results and business intent results.

[0015] The physical layer intent result is obtained based on the network intent result, and the business layer intent result is obtained based on the business intent result.

[0016] The intent recognition result is obtained by combining the intent results from the physical layer and the intent results from the business layer.

[0017] In one embodiment, joint semantic enhancement is performed on data from all modalities, including:

[0018] Popularity construction is performed on multimodal data using a self-evolving semantic module to obtain popularity construction results; the popularity construction results characterize the distribution of feature vectors of multimodal data.

[0019] The popular construction results are decomposed into feature values ​​to obtain the feature value vector of the popular construction results. The feature value vector is then distilled based on a preset feature value threshold to obtain the joint semantic enhancement result.

[0020] In one embodiment, attention fusion is performed on features of a first type of data from at least two modalities, including:

[0021] The attention matrix is ​​obtained by performing Hadamard product processing on the features of the first type of data in at least two modalities using the quantum fusion module.

[0022] The attention matrix is ​​applied to any modality of the first type of data in the multimodal features to obtain the feature fusion result of the multimodal features.

[0023] In one embodiment, the attention fusion result, the joint semantic enhancement result, and the features of the second type of data from at least two modalities are dynamically balanced to obtain the target data result, including:

[0024] The attention fusion result, the joint semantic enhancement result, and the features of the second type of data from at least two modalities are input into the gated fusion module. Based on the gated fusion module, the features of the attention fusion result, the joint semantic enhancement result, and the features of the second type of data from at least two modalities are summed element-wise to obtain a comprehensive feature vector.

[0025] The comprehensive feature vector is nonlinearly activated based on a preset logic function to obtain the fusion weights. The fusion weights are then combined with the comprehensive feature vectors to obtain the target data result.

[0026] In one embodiment, obtaining the physical layer intent result based on the network intent result and obtaining the service layer intent result based on the service intent result includes:

[0027] All feature vectors included in the network intent result are identified, and the feature vectors are converted into probabilities. The network intent corresponding to the feature vector with the highest probability is determined as the physical layer intent result. Among them, there is a one-to-one correspondence between the element index of the feature vector and the preset network intent.

[0028] All feature vectors included in the business intent result are identified, and the feature vectors are converted into probabilities. The business intents corresponding to feature vectors with probabilities greater than a preset probability threshold are determined as the business layer intent results. The element indices of the feature vectors correspond one-to-one with the preset business intents.

[0029] In one embodiment, the multimodal data includes text data, topology graph data, and traffic sequence data; at least two modalities of data are input into a pre-trained intent recognition model, joint semantic enhancement is performed on all modalities of data, attention fusion is performed on the features of the first type of data in the at least two modalities of data, and the attention fusion result, the joint semantic enhancement result, and the features of the second type of data in the at least two modalities of data are dynamically balanced to obtain the target data result, including:

[0030] Joint semantic enhancement is performed on text data, topology graph data, and traffic sequence data. Attention fusion is performed on the features of text data and topology graph data. The attention fusion result, joint semantic enhancement result, and features of traffic sequence data are dynamically balanced to obtain the target data result.

[0031] In one embodiment, obtaining a fully trained intent recognition model includes:

[0032] Obtain the initial intent recognition model and the preset training set, wherein the training set includes multimodal training data and labels corresponding to the training data of each modality, and the labels include network intent labels and business intent labels;

[0033] The training set is input into the initial intent recognition model for training to obtain the training prediction result. The loss function result is calculated based on the training prediction result and the label. The gradient of the loss function result is backpropagated to the initial intent recognition model for iterative training to generate a fully trained intent recognition model.

[0034] Secondly, this application also provides a network planning intent recognition device, comprising:

[0035] The acquisition module is used to acquire data from at least two modalities of the network to be identified;

[0036] The computation module is used to input data from at least two modalities into a pre-trained intent recognition model, perform joint semantic enhancement on the data from all modalities, perform attention fusion on the features of the first type of data in the data from at least two modalities, and perform dynamic balancing processing on the attention fusion result, the joint semantic enhancement result, and the features of the second type of data in the data from at least two modalities to obtain the target data result; wherein, the data from the at least two modalities are divided according to a preset partitioning rule to obtain the first type of data and the second type of data;

[0037] The generation module is used to identify the target data results and obtain the network planning intent identification results corresponding to the network to be identified.

[0038] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0039] Acquire data for at least two modalities of the network to be identified;

[0040] At least two modalities of data are input into a pre-trained intent recognition model. Joint semantic enhancement is performed on all modalities of data. Attention fusion is performed on the features of the first type of data in the at least two modalities of data. The attention fusion result, the joint semantic enhancement result, and the features of the second type of data in the at least two modalities of data are dynamically balanced to obtain the target data result. Among them, the at least two modalities of data are divided according to a preset partitioning rule to obtain the first type of data and the second type of data.

[0041] The target data results are identified to obtain the network planning intent identification results corresponding to the network to be identified.

[0042] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0043] Acquire data for at least two modalities of the network to be identified;

[0044] At least two modalities of data are input into a pre-trained intent recognition model. Joint semantic enhancement is performed on all modalities of data. Attention fusion is performed on the features of the first type of data in the at least two modalities of data. The attention fusion result, the joint semantic enhancement result, and the features of the second type of data in the at least two modalities of data are dynamically balanced to obtain the target data result. Among them, the at least two modalities of data are divided according to a preset partitioning rule to obtain the first type of data and the second type of data.

[0045] The target data results are identified to obtain the network planning intent identification results corresponding to the network to be identified.

[0046] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0047] Acquire data for at least two modalities of the network to be identified;

[0048] At least two modalities of data are input into a pre-trained intent recognition model. Joint semantic enhancement is performed on all modalities of data. Attention fusion is performed on the features of the first type of data in the at least two modalities of data. The attention fusion result, the joint semantic enhancement result, and the features of the second type of data in the at least two modalities of data are dynamically balanced to obtain the target data result. Among them, the at least two modalities of data are divided according to a preset partitioning rule to obtain the first type of data and the second type of data.

[0049] The target data results are identified to obtain the network planning intent identification results corresponding to the network to be identified.

[0050] The aforementioned method, apparatus, device, medium, and product for network planning intent recognition include inputting acquired data from at least two modalities into an intent recognition model; performing joint semantic enhancement on the data from all modalities; performing attention fusion on the features of a first type of data from the at least two modalities; and performing dynamic balancing processing on the attention fusion result, the joint semantic enhancement result, and the second type of data from the at least two modalities to obtain a target data result. Specifically, the data from the at least two modalities are divided based on a preset partitioning rule to obtain first type data and second type data. Finally, the target data result is recognized to obtain a network planning intent recognition result corresponding to the network to be recognized. In summary, this application first performs semantic enhancement on the data through joint semantic enhancement computation, removing noise and redundant information. Then, it performs attention fusion on the features of the first type of data to establish the association between data of different modalities. Finally, it performs dynamic balancing processing on the attention fusion result, the joint semantic enhancement result, and the features of the second type of data, which can dynamically adjust the weights of data of different modalities. For example, if a certain modality has high noise or weak information, it will be suppressed through weighting, thereby enhancing the robustness of the model to imbalanced modalities, thus obtaining more accurate target data results, and improving the accuracy of the final planning intent recognition result corresponding to the network to be identified. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is an application environment diagram of the network planning intent recognition method in one embodiment;

[0053] Figure 2 This is a flowchart illustrating a network planning intent recognition method in one embodiment;

[0054] Figure 3 This is a schematic diagram of the structure of an intent recognition model in one embodiment;

[0055] Figure 4 This is a schematic diagram of the process for generating network planning intent recognition results in one embodiment;

[0056] Figure 5 This is a flowchart illustrating the attention fusion process for features of a first type of data in one embodiment;

[0057] Figure 6This is a flowchart illustrating the process of generating target data results in one embodiment;

[0058] Figure 7 This is a flowchart illustrating the process of generating physical layer intent results and business layer intent results in one embodiment.

[0059] Figure 8 This is a schematic diagram of the structure of an intent recognition model in one embodiment;

[0060] Figure 9 This is a structural block diagram of a network planning intent recognition device in one embodiment. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0062] The technical background of this application will be explained below.

[0063] Network planning intent identification refers to the intelligent analysis process of extracting network optimization objectives (such as bandwidth adjustment and disaster recovery backup) from multimodal and multi-source data. Network planning intent identification is typically based on feature analysis of data across multiple network modalities, which generally reflect different aspects of the network's characteristics. However, in related technologies for network planning based on multimodal data, the data structures and types of data from different modalities vary significantly. Furthermore, many technologies process multimodal data independently, neglecting the inter-modal relationships. Therefore, the data analysis results are generally poor, failing to effectively correlate and identify different modalities. This leads to an inability to accurately identify network planning intent, and consequently, an inability to accurately adjust and optimize the network.

[0064] Based on this, this application provides a network planning intent recognition method. The method inputs acquired data from multiple modalities into a pre-trained intent recognition model. First, joint semantic enhancement is performed on the data from all modalities. Then, attention fusion is performed on the features of non-temporal data from multiple modalities. The attention fusion result, the joint semantic enhancement result, and the temporal data from multiple modalities are dynamically balanced to obtain the target data result. This application can effectively associate and recognize data from different modalities, and maintain high recognition accuracy even when some modal data is incomplete. See the following embodiments for details:

[0065] The network planning intent recognition method provided in this application embodiment can be applied to, for example, Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated on server 104 or placed on a cloud or other network server. First, data from at least two modalities of the network to be identified is acquired. Then, the data from at least two modalities is input into a pre-trained intent recognition model. Joint semantic enhancement is performed on the data from all modalities. Attention fusion is performed on the features of non-temporal data from the at least two modalities. The attention fusion result, the joint semantic enhancement result, and the features of temporal data from the at least two modalities are dynamically balanced to obtain the target data result. Finally, the target data result is recognized to obtain the corresponding network planning intent recognition result. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can be smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, projection devices, etc. Portable wearable devices can be smartwatches, smart bracelets, head-mounted devices, etc. Headset devices can be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. Server 104 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server providing cloud computing services.

[0066] In one exemplary embodiment, such as Figure 2 As shown, a method for identifying network planning intent is provided, which can be applied to... Figure 1 Taking the server in the example of this, the explanation includes:

[0067] Step S210: Obtain data for at least two modalities of the network to be identified.

[0068] The network to be identified can be any network requiring network planning, including but not limited to the Internet, the Internet of Things, and local area networks. This network can be applied in various scenarios, such as hospital networks and industrial networks. The data of at least two modalities of the network to be identified includes, but is not limited to, network traffic data, text data related to the network (e.g., in a hospital network scenario, this text data could be "ensuring the real-time performance of remote surgery" input by relevant technicians), and a network topology diagram (again, using a hospital network as an example, this topology diagram could be a connection diagram of network-controlled medical devices within the hospital). This embodiment does not impose excessive restrictions on the number of modalities; two or more modalities are sufficient. Similarly, this embodiment does not impose excessive restrictions on the specific type of modality; text data, topology diagrams, traffic data, and other modal types are all acceptable.

[0069] Step S220: Input data from at least two modalities into a pre-trained intent recognition model, perform joint semantic enhancement on all modalities, perform attention fusion on the features of the first type of data in the at least two modalities, and perform dynamic balancing on the attention fusion result, the joint semantic enhancement result, and the features of the second type of data in the at least two modalities to obtain the target data result. The at least two modalities are divided based on a preset partitioning rule to obtain the first type of data and the second type of data.

[0070] In this embodiment, data from at least two modalities are input into a pre-trained intent recognition model. The intent recognition model is a neural network model established in this application, which includes various preset modules, such as feature extraction modules, semantic enhancement modules, etc. Those skilled in the art can add or remove modules in the intent recognition model according to actual needs.

[0071] Data from at least two modalities is input into the aforementioned intent recognition model. First, joint semantic enhancement is performed on the data from all modalities through a preset module in the intent recognition model. Specifically, joint semantic enhancement can be performed on the data from all modalities according to the semantic enhancement module in the intent recognition model. The joint semantic enhancement method includes, but is not limited to, first using manifold learning and eigenvalue decomposition to reduce the dimensionality of the features and extract principal components, and then setting the eigenvalues ​​of the feature vectors below a preset threshold to zero, thereby achieving feature distillation, removing noise and redundant information, and thus obtaining the semantically enhanced feature vectors.

[0072] Then, attention fusion is performed on the features of the first type of data in at least two modalities through another preset module in the intent recognition model. In one embodiment, the attention fusion method includes, but is not limited to, fine-grained fusion based on cross-modal attention. Taking image data and text data as an example, the attention fusion method includes, but is not limited to, first extracting features from the data of different modalities, assigning weights to the feature vectors of any modality (e.g., assigning weights to image feature vectors), and then fusing the weighted image feature vectors with the text feature vectors to obtain the attention fusion result. In another embodiment, quantum attention fusion can be performed on the text feature vectors and image feature vectors, that is, simulating the quantum entanglement state to calculate the correlation between text and image, and performing weighted output based on the correlation. Through attention fusion, physical correlation between data of different modalities can be established. Preferably, quantum attention fusion can be used to replace traditional dot product attention calculation.

[0073] Finally, another pre-defined module in the intent recognition model dynamically balances the features of the attention fusion result, the joint semantic enhancement result, and the second type of data in at least two modalities. In one embodiment, this includes dynamically assigning weights to the data of different modalities based on the data quality of each modality. For example, the features of different modalities are mapped to the same semantic space, and weights are generated by a gating network based on the content (i.e., data quality) of each modal feature. The weights are then used to fuse the features of each modality.

[0074] In summary, the target data results after semantic enhancement, attention fusion, and modality dynamic balancing can be obtained through the above methods. In practical applications, if there are other modules that can perform semantic enhancement, attention fusion, or dynamic balancing, they can also be introduced into the intent recognition model. The modules included in the intent recognition model in this application embodiment can be deleted or added according to actual needs.

[0075] It should be further noted that this application has multiple methods for classifying data with at least two modalities, including but not limited to, in one embodiment, data can be classified according to data type. For example, data with at least two modalities can be classified into time-series data and non-time-series data. For instance, when the data modalities include traffic time-series data, text data, and topology data, traffic time-series data can be classified as the first type of data, and text data and topology data can be classified together as the second type of data. Alternatively, text data and topology data can be classified together as the first type of data, and traffic time-series data can be classified as the second type of data. That is, this classification method focuses on the type of data itself.

[0076] In another embodiment, data can be divided into physical layer and intent layer. The physical layer refers to the basic data used to describe the network, such as topology data (e.g., in a hospital scenario, the topology data could be a connection diagram of medical equipment in the hospital area). The intent layer refers to data that directly reflects business needs, such as text data (e.g., in a hospital scenario, the text data could be: ensuring the real-time performance of remote surgery). In this case, assuming that the data modality includes traffic time-series data (e.g., in a hospital scenario, traffic time-series data can reflect network bandwidth, connectivity, etc.), text data, and topology data, traffic time-series data and topology data can be classified as the first type of data, and text data can be classified as the second type of data. Alternatively, traffic time-series data and topology data can be classified as the second type of data, and text data can be classified as the first type of data. That is, this classification method focuses on the intent type contributed by the essence of the data.

[0077] In summary, by dividing data of at least two modalities into first-type data and second-type data according to different classification methods, data of the same type can be better integrated.

[0078] It is understood that the above data types are not limited to topology data, text data, and traffic time-series data, and the partitioning rules are not limited to the above two partitioning rules. Those skilled in the art can add or reduce the data types to be calculated according to actual needs, and can also use other partitioning rules according to actual situations.

[0079] Step S230: Identify the target data results to obtain the network planning intent identification results corresponding to the network to be identified.

[0080] In this embodiment of the application, the target data result is identified by the identification module in the intent recognition model to obtain the above-mentioned network planning intent recognition result. The network planning intent recognition result represents the network planning scheme obtained based on multimodal data. Executing the planning scheme (i.e., executing the network planning intent recognition result) can effectively adjust the network to make it more in line with the expectations of multimodal data.

[0081] In summary, joint semantic enhancement can be performed on multimodal data to filter out noise or minor details, achieving semantic space compression and evolution. Furthermore, attention fusion is applied to the features of the first type of data from at least two modalities, allowing direct interaction between multimodal features and capturing local correlations. This fusion method is more refined than simple feature concatenation and enhances semantic alignment between modalities. Finally, another pre-defined module dynamically balances the attention fusion result, the joint semantic enhancement result, and the features of the second type of data from at least two modalities. This dynamically adjusts the weights of different modalities; for example, if a modality has high noise or weak information, it will be suppressed through weighting, thereby enhancing the model's robustness to imbalanced modalities. Therefore, through the embodiments of this application, more accurate network planning intent recognition results can be obtained.

[0082] In an exemplary embodiment, the intent recognition model includes a self-evolving semantic module, a quantum fusion module, and a gated fusion module. Data from at least two modalities is input into the pre-trained intent recognition model. Joint semantic enhancement is performed on the data from all modalities. Attention fusion is performed on the features of the first type of data from the at least two modalities. The attention fusion result, the joint semantic enhancement result, and the features of the second type of data from the at least two modalities are dynamically balanced to obtain the target data result, including:

[0083] Joint semantic enhancement is performed on data from all modalities through a self-evolving semantic module;

[0084] Attention fusion of features of first-type data from at least two modalities is performed using a quantum fusion module.

[0085] The gating fusion module dynamically balances the attention fusion result, the joint semantic enhancement result, and the features of the second type of data from at least two modalities to obtain the target data result.

[0086] In this embodiment of the application, the above-mentioned intent recognition model includes a self-evolving semantic module, a quantum fusion module, and a gated fusion module. It can also be understood that the intent recognition model further includes a feature extraction module corresponding to the data of each modality. Figure 3This is a schematic diagram of the intent recognition model in one embodiment. As shown in the diagram, this embodiment uses three modalities of data as an example. In practical applications, fewer or more modalities can be used, depending on actual needs. The diagram shows that modality 1 data, modality 2 data, and modality 3 data are all input into the self-evolving semantic module. The self-evolving semantic module performs joint semantic enhancement on all modalities of data. Modality 1 data and modality 2 data are respectively processed by the feature extraction module and then input into the quantum fusion module for attention fusion. The quantum fusion module simulates quantum entanglement through matrix operations, establishing a physical connection between modality 1 data and modality 2 data, replacing traditional dot product attention calculation. Modality 1 data and modality 2 data are both set as first-type data. Finally, the attention fusion result, the joint semantic enhancement result, and the features of modality 3 data (i.e., second-type data) are input into the gating fusion module for dynamic balancing to obtain the aforementioned target data result.

[0087] In one exemplary embodiment, such as Figure 4 As shown, the intent recognition model also includes an intent decoupling module. The process of recognizing the target data results to obtain the network planning intent recognition result corresponding to the network to be recognized includes:

[0088] Step S410: The intent decoupling module is used to segment the preset dimension region in the target data result to obtain the network intent result and the business intent result.

[0089] In this embodiment, the target data result is hard-segmented by the intent decoupling module to distinguish between network intent result and service intent result. The network intent result represents device-oriented, technical-level adjustments to the network, such as bandwidth adjustment and path switching, while the service intent result represents business-oriented, more abstract, higher-level service intent, such as activating a zero-jitter strategy. Combining the network intent result and the service intent result can better achieve intelligent network management.

[0090] The hard segmentation scheme includes, but is not limited to, hard segmenting the first multi-dimensional data of the target data result and using the segmented first multi-dimensional data as the network intent result. Similarly, hard segmenting the last multi-dimensional data of the target data result and using the segmented last multi-dimensional data as the business intent result. In one embodiment, the target data result calculated above is generally 128-dimensional. In this case, the first 4 dimensions can be used as the network intent result and the last 4 dimensions can be used as the business intent result.

[0091] The above solution can be implemented using hard index partitioning. The following shows some of the main code:

[0092] INDArray baseIntent = fused.getColumns(0, 3);

[0093] INDArray bizIntent = fused.getColumns(4, 7);

[0094] Step S420: Obtain the physical layer intent result based on the network intent result, and obtain the service layer intent result based on the service intent result.

[0095] In this embodiment, the network intent result is a 4-dimensional vector, such as [z1, z2, z3, z4]. These four-dimensional vectors are converted into probabilities using the Softmax function, and the probabilities are normalized so that the sum of the four probabilities is 1. Thus, a unique physical layer intent result is determined based on the value with the highest probability. Similarly, the service intent result is also a 4-dimensional vector [y1, y2, y3, y4]. These four-dimensional vectors are converted into probabilities using the Sigmoid activation function, and at least one service layer intent result is determined based on the probability exceeding a preset threshold (e.g., 0.7).

[0096] Step S430: Combine the intent results from the physical layer and the intent results from the business layer to obtain the intent recognition result.

[0097] Both the intent results from the physical layer and the intent results from the business layer are used as the intent recognition results mentioned above.

[0098] In one exemplary embodiment, joint semantic enhancement is performed on data from all modalities, including:

[0099] Popularity construction is performed on multimodal data using a self-evolving semantic module to obtain popularity construction results; the popularity construction results characterize the distribution of feature vectors of multimodal data.

[0100] The popular construction results are decomposed into feature values ​​to obtain the feature value vector of the popular construction results. The feature value vector is then distilled based on a preset feature value threshold to obtain the joint semantic enhancement result.

[0101] This application embodiment defines a joint semantic enhancement scheme, which is implemented based on the self-evolving semantic module in the intent recognition model.

[0102] First, all modal data are input into the self-evolving semantic module. The self-evolving semantic module uses manifold learning to reduce the dimensionality and reconstruct the input features. This includes capturing the distribution characteristics of the data by calculating structures such as the covariance matrix of the data, which represents the correlation between various features. This results in a manifold construction result that can characterize the distribution of feature vectors of multimodal data. The manifold construction result is a covariance matrix.

[0103] The popular construction results are then subjected to eigenvalue decomposition, generating eigenvalues ​​and eigenvectors. Feature distillation is performed on the eigenvectors based on a preset feature threshold, which can be set to 0.75 times the average value of the eigenvalues, or preset by relevant technical personnel according to actual needs. Each eigenvector is then iterated through, comparing each eigenvalue with the threshold. Only eigenvectors with eigenvalues ​​greater than the threshold are retained, while those below the threshold are set to zero, achieving semantic space compression and purification, resulting in more accurate recognition results. Finally, the updated eigenvectors are output, representing the joint semantic enhancement result. This joint semantic enhancement result can also update the semantic space of the self-evolving semantic module itself, enabling the model to adapt to new data without complete retraining, ensuring the model adapts to dynamic semantic environments and facilitating long-term model evolution.

[0104] In one exemplary embodiment, such as Figure 5 As shown, attention fusion is performed on the features of the first type of data in at least two modalities, including:

[0105] Step S510: The features of the first type of data in at least two modalities are processed by the Hadamard product through the quantum fusion module to obtain the attention matrix.

[0106] In this embodiment, a preset feature extraction module first extracts features from the first type of data to obtain the features of the first type of data. Then, the features of the first type of data are input into the aforementioned quantum fusion module. The quantum fusion module performs element-wise multiplication on the features of the first type of data. This Hadamard multiplication simulates a quantum entanglement state, that is, it fuses the features of the first type of data element-wise, highlighting the common features and weakening the inconsistent ones. The Hadamard multiplication result is then subjected to probability normalization, achieved by dividing the L2 norm by the dimension, thus obtaining the aforementioned attention matrix. In this embodiment, the attention matrix serves as the attention weights.

[0107] Step S520: Apply the attention matrix to any one of the data features of the first type of data in the multimodal features to obtain the feature fusion result of the multimodal features.

[0108] In this embodiment, the attention matrix described above is applied to any modality of the first type of data features input into the quantum fusion module. This can be achieved through element-wise multiplication, thereby scaling the elements in any modality of the first type of data features according to their weights. That is, elements with larger weights are strengthened, and elements with smaller weights are suppressed. This strengthens the relevant parts of the first type of data features input into the quantum fusion module, making the final feature fusion result more focused on the relevant semantics. It is understood that the quantum fusion module in this embodiment is applicable to combinations of multiple modalities.

[0109] The following shows a portion of the main code for the quantum fusion module:

[0110] INDArray entanglement = text.mul(graph);

[0111] INDArray prob = entanglement.norm2(1).div(entanglement.shape()[1]);

[0112] return text.mul(prob.reshape(prob.length(), 1)).

[0113] In one exemplary embodiment, such as Figure 6 As shown, the attention fusion result, the joint semantic enhancement result, and the features of the second type of data from at least two modalities are dynamically balanced to obtain the target data result, including:

[0114] Step S610: Input the attention fusion result, the joint semantic enhancement result, and the features of the second type of data in at least two modalities into the gated fusion module. Based on the gated fusion module, perform element-wise summation on the features of the attention fusion result, the joint semantic enhancement result, and the features of the second type of data in at least two modalities to obtain a comprehensive feature vector.

[0115] In this embodiment of the application, the gating fusion module can adaptively adjust data from different modalities, specifically:

[0116] First, the obtained attention fusion results, joint semantic enhancement results, and features of the second type of data are input into the gating fusion module. The gating fusion module first performs element-wise summation on the attention fusion results, joint semantic enhancement results, and features of the second type of data, that is, directly adding the data at corresponding positions, which is equivalent to initially integrating the information of different modalities to obtain the above comprehensive feature vector.

[0117] Step S620: Based on a preset logic function, the comprehensive feature vector is nonlinearly activated to obtain the fusion weights, and the fusion weights are combined with the comprehensive feature vectors to obtain the target data result.

[0118] In this embodiment, a preset logic function is applied to each element in the comprehensive feature vector. This logic function can be a Sigmoid function. The comprehensive feature vector is then non-linearly activated based on the logic function to obtain the weight of each element. The value range is generally (0,1). Based on the weight of each element in the comprehensive feature vector, the above-mentioned fusion weight is obtained.

[0119] Finally, the fusion weights and comprehensive feature vectors are combined to obtain the target data result. In practical applications, this process is data-driven and can adaptively adjust the contribution weights of each modality (for example, multimodal data includes traffic second-type data; this application can automatically increase the overall influence of traffic data when traffic anomalies occur), without explicitly assigning independent weights or performing weighted summation; the target data result can be directly output. The following shows some of the main code of the gating fusion module:

[0120] INDArray gate = Nd4j.ones(modalities[0].shape());

[0121] for(INDArray mod : modalities) {

[0122] gate = gate.add(mod);

[0123] }

[0124] INDArray gateSig = Sigmoid(gate).

[0125] Through the embodiments of this application, the attention fusion result, the joint semantic enhancement result, and the second type of data in at least two modal data are dynamically balanced. The comprehensive feature vector obtained by summation can be nonlinearly activated to obtain the contribution weight of each modality. Thus, the weight of each modality can be determined according to the data quality of each data type (in the embodiments of this application, the data quality is determined by the numerical value of the feature itself), thereby dynamically fusing the input data of each type and reducing the impact of abnormal data.

[0126] In one exemplary embodiment, such as Figure 7 As shown, the physical layer intent result is obtained based on the network intent result, and the business layer intent result is obtained based on the business intent result, including:

[0127] Step S710: Determine all feature vectors included in the network intent result, convert the feature vectors into probabilities, and determine the network intent corresponding to the feature vector with the highest probability as the physical layer intent result; wherein, the element index of the feature vector corresponds one-to-one with the preset network intent.

[0128] As described above, the network intent result includes multiple feature vectors, and the data of the feature vectors is determined by the segmentation rules. For example, if the first four dimensions of the target data result are assigned to the network intent result, then the network intent result includes the first four-dimensional vector. Similarly, the feature vectors included in the business intent result are also determined by the preset segmentation rules. If the last four dimensions of the target data result are assigned to the business intent result, then the business intent result includes the last four-dimensional vector.

[0129] Taking the assignment of the first four features to the network intent result and the last four features to the business intent result as an example, the decoupling module in the intent recognition model cuts the first four features in the target data result to determine the network intent result, and cuts the last four features in the target data result to determine the business intent result.

[0130] Then, the network intent conversion module transforms the four-dimensional feature vector contained in the network intent result into four probability distributions using the Softmax activation function, with the sum of the four probability distributions being 1. The data with the highest probability value among the four probability distributions is then determined, and the physical layer intent result is determined based on the position of the highest probability value. It can be understood that different positions in the first four dimensions correspond to different physical layer intent results. For example, the first position in the first four dimensions corresponds to bandwidth adjustment; that is, if the probability value at the first position is the highest, the physical layer intent result is bandwidth adjustment. Similarly, the second position in the first four dimensions corresponds to path switching; that is, if the probability value at the second position is the highest, the physical layer intent result is path switching, and so on. The element index of the feature vector represents the position of the feature vector.

[0131] Step S720: Determine all feature vectors included in the business intent result, convert the feature vectors into probabilities, and determine the business intent corresponding to the feature vector with a probability value greater than a preset probability threshold as the intent result of the business layer. The element index of the feature vector corresponds one-to-one with the preset business intent.

[0132] Similarly, the decoupling module obtains the last four-dimensional feature vector, which is the aforementioned business intent result. Then, through the business intent conversion module, the four-dimensional feature vector contained in the business intent result is converted into four probability distributions using the Sigmoid function. Then, at least one feature value in the four probability distributions with a probability value greater than a preset probability threshold is determined. The intent result of the aforementioned business layer is determined based on the position of the at least one feature value greater than the preset probability threshold. It can be understood that different positions in the last four dimensions correspond to different business layer intent results. The element index of the feature vector corresponds one-to-one with the preset business intent, and the element index of the feature vector represents the position of the feature vector.

[0133] It should be noted that the reason why element indices can be mapped to preset intents (including business intents and network intents) is implemented during the model training phase. That is, during training, the output layer is explicitly divided into two parts: one part is used for physical layer intent result recognition, and the other part is used for business layer intent result recognition. During the training phase, a loss function is used to require that the most correct result be selected from multiple network intents in physical layer intent recognition, while multiple correct business intents can be selected simultaneously in business intent recognition. Furthermore, when preparing the training set, it is labeled according to a preset mapping relationship, such as the first dimension corresponding to restarting the device and the second dimension corresponding to switching paths. Thus, the model can learn the correspondence between position (i.e., element index) and preset intents during training.

[0134] In an exemplary embodiment, the multimodal data includes text data, topology graph data, and traffic sequence data. At least two modalities of data are input into a pre-trained intent recognition model. Joint semantic enhancement is performed on all modalities of data. Attention fusion is performed on the features of the first type of data in the at least two modalities of data. The attention fusion result, the joint semantic enhancement result, and the features of the second type of data in the at least two modalities of data are dynamically balanced to obtain the target data result, including:

[0135] Joint semantic enhancement is performed on text data, topology graph data, and traffic sequence data. Attention fusion is performed on the features of text data and topology graph data. The attention fusion result, joint semantic enhancement result, and features of traffic sequence data are dynamically balanced to obtain the target data result.

[0136] This application provides a preferred embodiment, such as... Figure 8 As shown, multimodal data, including text data, topology graph data, and traffic sequence data, is input into the aforementioned intent recognition model. First, a self-evolving semantic module performs joint semantic enhancement on all modalities. Then, pre-defined text feature extraction and image feature extraction modules extract features from the text and topology graph data, respectively. The extracted text and topology graph features are input into a quantum fusion module for attention fusion. Finally, the attention fusion result, the joint semantic enhancement result, and the features from the traffic sequence data are input into a gating fusion module. Based on the gating fusion module, the input data is dynamically balanced to obtain the target data recognition result. This target data recognition result is then input into a decoupling module, which segments it into network intent results and business intent results. A network intent conversion module identifies the network intent results to obtain the physical layer intent results. Similarly, a business intent conversion module identifies the business intent results to obtain the business layer intent results. The physical layer intent results and the business layer intent results together constitute the network planning intent recognition result corresponding to the network to be identified.

[0137] In one exemplary embodiment, obtaining a fully trained intent recognition model includes:

[0138] Obtain the initial intent recognition model and the preset training set, wherein the training set includes multimodal training data and labels corresponding to the training data of each modality, and the labels include network intent labels and business intent labels;

[0139] The training set is input into the initial intent recognition model for training to obtain the training prediction result. The loss function result is calculated based on the training prediction result and the label. The gradient of the loss function result is backpropagated to the initial intent recognition model for iterative training to generate a fully trained intent recognition model.

[0140] This application embodiment illustrates how to train the intent recognition model. The training set includes multimodal training data, the modality of which should be consistent with the modality of the input data in the actual application stage. The training data needs to be labeled, which can be a network intent label or a business intent label. The label is in the form of a multidimensional vector. For example, if the handling of a network fault is to adjust the bandwidth, the label should be set to [0,1,0,0], that is, the second position corresponds to "adjust bandwidth". The setting of the business intent label is similar, and the dimension of the label can be set by relevant technical personnel according to actual needs.

[0141] This application also provides a preferred embodiment of a network planning intent recognition method.

[0142] Taking the application of the solution in this application to a smart hospital network as an example, assume the input data includes:

[0143] Text data (text data may include phrases such as "ensure the real-time nature of remote surgery");

[0144] Topology data (topology data may include, for example, "connection diagram of medical equipment in the hospital area");

[0145] Traffic sequence data (which may include, for example, "surgical video stream burst data").

[0146] The aforementioned text data, topology data, and traffic sequence data are then input into the fully trained intent recognition model. The model uses a quantum fusion module to associate "real-time performance" with the endoscope device node, and a gating fusion module to increase the burst traffic weight to 0.96. The final output is assumed to be the network intent: [0.91, 0.04, 0.03, 0.02]. The first position has the highest probability value, triggering the corresponding network layer intent, i.e., a 10 Gigabit port upgrade. Similarly, for the service intent, at least one result with a probability value greater than a probability threshold is selected, and the service layer intent recognition result is determined based on the position of the probability value greater than the threshold, such as activating the "zero jitter transmission" strategy.

[0147] Experiments showed that the final surgical video stream latency fluctuation was less than 0.1ms.

[0148] It should be noted that the solution in this application can also be applied to industrial scenarios, operation and maintenance scenarios, etc.

[0149] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0150] Based on the same inventive concept, this application also provides a network planning intent recognition device for implementing the network planning intent recognition method described above. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations in one or more network planning intent recognition device embodiments provided below can be found in the limitations of the network planning intent recognition method described above, and will not be repeated here.

[0151] In one exemplary embodiment, such as Figure 9 As shown, a network planning intent recognition device is provided, comprising: an acquisition module 91, a calculation module 92, and a generation module 93, wherein:

[0152] Acquisition module 91 is used to acquire data of at least two modalities of the network to be identified;

[0153] The calculation module 92 is used to input the data of the at least two modalities into a pre-trained intent recognition model, perform joint semantic enhancement on the data of all modalities, perform attention fusion on the features of the first type of data in the data of the at least two modalities, and perform dynamic balancing processing on the attention fusion result, the joint semantic enhancement result, and the features of the second type of data in the data of the at least two modalities to obtain the target data result; wherein, the data of the at least two modalities are divided according to a preset partitioning rule to obtain the first type of data and the second type of data;

[0154] The generation module 93 is used to identify the target data results and obtain the network planning intent identification result corresponding to the network to be identified.

[0155] In one exemplary embodiment, the computing module further includes a second semantic computing module, a second fusion computing module, and a second gating computing module:

[0156] The second semantic computing module is used to perform joint semantic enhancement on data of all modalities through the self-evolving semantic module;

[0157] The second fusion computing module is used to perform attention fusion on the features of the first type of data in the data of the at least two modalities through the quantum fusion module;

[0158] The second gating computation module is used to dynamically balance the attention fusion result, the joint semantic enhancement result, and the features of the second type of data in the data of at least two modalities through the gating fusion module to obtain the target data result.

[0159] In one exemplary embodiment, the generation module further includes a third decoupling module, a third generation module, and a third result module:

[0160] The third decoupling module is used to segment the preset dimension region in the target data result through the intent decoupling module to obtain the network intent result and the business intent result;

[0161] The third generation module is used to obtain the physical layer intent result based on the network intent result, and to obtain the service layer intent result based on the service intent result.

[0162] The third result module is used to combine the intent results of the physical layer and the intent results of the business layer to obtain the intent recognition result.

[0163] In one exemplary embodiment, the network planning intent recognition device further includes a fourth construction module and a fourth decomposition module:

[0164] The fourth construction module is used to perform popularization construction on the multimodal data through the self-evolving semantic module to obtain popularization construction results; the popularization construction results characterize the distribution of feature vectors of the multimodal data.

[0165] The fourth decomposition module is used to perform feature value decomposition on the popular construction result to obtain the feature value vector of the popular construction result, and to perform feature distillation on the feature value vector based on a preset feature value threshold to obtain the joint semantic enhancement result.

[0166] In one exemplary embodiment, the network planning intent recognition device further includes a fifth fusion module and a fifth generation module:

[0167] The fifth fusion module is used to perform Hadamard product processing on the features of the first type of data in the data of the at least two modalities through the quantum fusion module to obtain the attention matrix;

[0168] The fifth generation module is used to apply the attention matrix to any one of the data features of the first type of data in the multimodal features to obtain the feature fusion result of the multimodal features.

[0169] In one exemplary embodiment, the network planning intent recognition device further includes a sixth fusion module and a sixth generation module:

[0170] The sixth fusion module is used to input the attention fusion result, the joint semantic enhancement result, and the features of the second type of data in the data of at least two modalities into the gated fusion module, and perform element-wise summation on the features of the attention fusion result, the joint semantic enhancement result, and the features of the second type of data in the data of at least two modalities based on the gated fusion module to obtain a comprehensive feature vector;

[0171] The sixth generation module is used to perform nonlinear activation on the comprehensive feature vector based on a preset logic function to obtain fusion weights, and combine the fusion weights with the comprehensive feature vectors to obtain the target data result.

[0172] In one exemplary embodiment, the third generation module further includes a network intent conversion module and a service intent conversion module:

[0173] The network intent conversion module is used to determine all feature vectors included in the network intent result, convert the feature vectors into probabilities, and determine the network intent corresponding to the feature vector with the highest probability as the intent result of the physical layer; wherein, there is a one-to-one correspondence between the element index of the feature vector and the preset network intent;

[0174] The business intent conversion module is used to determine all feature vectors included in the business intent result, convert the feature vectors into probabilities, and determine the business intent corresponding to the feature vector with a probability value greater than a preset probability threshold as the intent result of the business layer, wherein there is a one-to-one correspondence between the element index of the feature vector and the preset business intent.

[0175] In one exemplary embodiment, the network planning intent recognition device further includes an eighth computing module:

[0176] The eighth calculation module is used to perform joint semantic enhancement on the text data, the topology graph data, and the traffic sequence data, perform attention fusion on the features of the text data and the topology graph data, and perform dynamic balancing processing on the attention fusion result, the joint semantic enhancement result, and the features of the traffic sequence data to obtain the target data result.

[0177] In one exemplary embodiment, the network planning intent recognition device further includes a ninth acquisition module and a ninth training module;

[0178] The ninth acquisition module is used to acquire the initial intent recognition model and the preset training set, wherein the training set includes multimodal training data and labels corresponding to the training data of each modality, and the labels include network intent labels and business intent labels.

[0179] The ninth training module is used to input the training set into the initial intent recognition model for training, obtain training prediction results, calculate the loss function result based on the training prediction result and the label, and backpropagate the gradient of the loss function result to the initial intent recognition model for iterative training to generate the fully trained intent recognition model.

[0180] The modules in the aforementioned network planning intent recognition device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0181] In one exemplary embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program that, when executed by the processor, implements any of the network planning intent recognition methods described above.

[0182] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements any of the network planning intent recognition methods described above.

[0183] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements any of the network planning intent recognition methods described above.

[0184] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0185] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0186] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for identifying network planning intent, characterized in that, The method includes: Acquire data for at least two modalities of the network to be identified; The data from at least two modalities are input into a pre-trained intent recognition model. Joint semantic enhancement is performed on the data from all modalities. Attention fusion is performed on the features of the first type of data in the data from at least two modalities. The attention fusion result, the joint semantic enhancement result, and the features of the second type of data in the data from at least two modalities are dynamically balanced to obtain the target data result. The data from at least two modalities are divided according to a preset partitioning rule to obtain the first type of data and the second type of data. The target data results are identified to obtain the network planning intent identification result corresponding to the network to be identified.

2. The method according to claim 1, characterized in that, The intent recognition model includes a self-evolving semantic module, a quantum fusion module, and a gated fusion module. The process involves inputting data from at least two modalities into the pre-trained intent recognition model, performing joint semantic enhancement on all modalities, performing attention fusion on the features of the first type of data in the at least two modalities, and dynamically balancing the attention fusion result, the joint semantic enhancement result, and the features of the second type of data in the at least two modalities to obtain the target data result, including: The self-evolving semantic module performs joint semantic enhancement on data from all modalities. The quantum fusion module performs attention fusion on the features of the first type of data in the data of at least two modalities. The gating fusion module dynamically balances the attention fusion result, the joint semantic enhancement result, and the features of the second type of data in the data of at least two modalities to obtain the target data result.

3. The method according to claim 1, characterized in that, The intent recognition model further includes an intent decoupling module. The step of recognizing the target data results to obtain the network planning intent recognition result corresponding to the network to be identified includes: The intent decoupling module segments the target data result into a preset dimension region to obtain network intent result and business intent result. Based on the network intent result, the physical layer intent result is obtained; based on the service intent result, the service layer intent result is obtained. The intent recognition result is obtained by combining the intent results from the physical layer and the intent results from the business layer.

4. The method according to any one of claims 1 to 3, characterized in that, The joint semantic enhancement of data across all modalities includes: Popularity construction is performed on all modalities of data using a self-evolving semantic module to obtain popularity construction results; the popularity construction results characterize the distribution of feature vectors of the multimodal data; The popular construction result is decomposed into feature value vectors, and the feature value vectors are distilled based on a preset feature value threshold to obtain joint semantic enhancement results.

5. The method according to any one of claims 1 to 3, characterized in that, The attention fusion of features of the first type of data in the at least two modalities includes: The attention matrix is ​​obtained by performing Hadamard product processing on the features of the first type of data in the data of at least two modalities through the quantum fusion module. The attention matrix is ​​applied to the data features of any one modality in the first type of data to obtain the feature fusion result of the multimodal features.

6. The method according to any one of claims 1 to 3, characterized in that, The dynamic balancing of the attention fusion result, the joint semantic enhancement result, and the features of the second type of data from the at least two modalities to obtain the target data result includes: The attention fusion result, the joint semantic enhancement result, and the features of the second type of data in the data of at least two modalities are input into the gated fusion module. Based on the features of the attention fusion result, the joint semantic enhancement result, and the features of the second type of data in the data of at least two modalities, the gated fusion module performs element-wise summation on the features of the attention fusion result, the joint semantic enhancement result, and the features of the second type of data in the data of at least two modalities to obtain a comprehensive feature vector. The comprehensive feature vector is nonlinearly activated based on a preset logic function to obtain a fusion weight, and the fusion weight is combined with the comprehensive feature vector to obtain the target data result.

7. The method according to claim 3, characterized in that, The process of obtaining the physical layer intent result based on the network intent result and obtaining the service layer intent result based on the service intent result includes: All feature vectors included in the network intent result are determined, and the feature vectors are converted into probabilities. The network intent corresponding to the feature vector with the highest probability is determined as the intent result of the physical layer. The element index of the feature vector corresponds one-to-one with the preset network intent. All feature vectors included in the business intent result are determined, and the feature vectors are converted into probabilities. The business intents corresponding to feature vectors with probability values ​​greater than a preset probability threshold are determined as the intent results of the business layer. The element indices of the feature vectors correspond one-to-one with the preset business intents.

8. The method according to any one of claims 1 to 3, characterized in that, The data of at least two modalities includes text data, topology graph data, and traffic sequence data; the process involves inputting the data of at least two modalities into a pre-trained intent recognition model, performing joint semantic enhancement on the data of all modalities, performing attention fusion on the features of the first type of data in the data of at least two modalities, and dynamically balancing the attention fusion result, the joint semantic enhancement result, and the features of the second type of data in the data of at least two modalities to obtain the target data result, including: Joint semantic enhancement is performed on the text data, the topology graph data, and the traffic sequence data. Attention fusion is performed on the features of the text data and the topology graph data. The attention fusion result, the joint semantic enhancement result, and the features of the traffic sequence data are then dynamically balanced to obtain the target data result.

9. The method according to any one of claims 1 to 3, characterized in that, Obtaining the fully trained intent recognition model includes: Obtain an initial intent recognition model and a preset training set, wherein the training set includes multimodal training data and labels corresponding to the training data of each modality, and the labels include network intent labels and business intent labels; The training set is input into the initial intent recognition model for training to obtain training prediction results. The loss function result is calculated based on the training prediction result and the label, and the gradient of the loss function result is backpropagated to the initial intent recognition model for iterative training to generate the fully trained intent recognition model.

10. A network planning intent recognition device, characterized in that, The device includes: The acquisition module is used to acquire data from at least two modalities of the network to be identified; The computation module is used to input the data from the at least two modalities into a pre-trained intent recognition model, perform joint semantic enhancement on the data from all modalities, perform attention fusion on the features of the first type of data in the at least two modalities, and perform dynamic balancing processing on the attention fusion result, the joint semantic enhancement result, and the features of the second type of data in the at least two modalities to obtain the target data result; wherein, the data from the at least two modalities are divided according to a preset partitioning rule to obtain the first type of data and the second type of data; The generation module is used to identify the target data results and obtain the network planning intent identification result corresponding to the network to be identified.

11. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 9.

13. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 9.

Images