Voice interaction system and method based on artificial intelligence technology

Abstract

The invention discloses a voice interaction system and method based on an artificial intelligence technology, and relates to the technical field of voice interaction, and the method comprises the steps: employing the double-layer cooperative recognition of a main channel Transform model and an auxiliary channel CTC-Attention mixed architecture, obtaining a text with a multi-dimensional feature label, carrying out the scene classification and dependency syntactic analysis through a bidirectional recurrent neural network, and obtaining a multi-dimensional feature label text; retrieving associated entities in real time and generating structured intention data in combination with a dynamic knowledge graph, executing three-dimensional analysis through an interpretability engine and generating a multi-dimensional traceability report, and triggering an intelligent decision fusion engine through a confidence threshold to generate structured enhanced interaction data; according to the method, scene classification and dependency syntactic analysis are carried out through the bidirectional recurrent neural network, the associated entities are retrieved in real time in combination with the dynamic knowledge graph, and the structured intention data are generated, so that the intention of the user is accurately understood, and personalized interaction services are provided.

Description

Technical Field

[0001] The present invention relates to the field of voice interaction technology, and in particular to a voice interaction system and method based on artificial intelligence technology. Background Art

[0002] With the continuous advancement of artificial intelligence technology, voice interaction systems have achieved significant development and application in many fields. Traditional voice interaction systems mainly rely on speech recognition technology and natural language processing technology. By converting the user's voice commands into text and then performing semantic analysis on the text to achieve interactive functions, the development of these technologies has greatly improved the convenience and naturalness of human-computer interaction, allowing users to easily control devices and obtain information through voice commands to complete complex tasks.

[0003] Although some improvement measures have been implemented in existing technologies, with the increasing complexity of application scenarios and the diversification of user needs, traditional voice interaction systems have gradually exposed many shortcomings. In terms of multimodal interaction, although existing technologies attempt to combine visual and tactile feedback, they lack effective synchronization mechanisms and interpretability, and cannot achieve true multimodal fusion. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a voice interaction method based on artificial intelligence technology to solve the problem that although the existing technology attempts to combine visual and tactile feedback, it lacks an effective synchronization mechanism and explainability.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a voice interaction method based on artificial intelligence technology, which includes: capturing user voice data, synchronizing a sound source focusing algorithm to lock the target direction, and using a deep neural network noise reduction model to eliminate environmental noise, and outputting a voice data packet;

[0008] It uses a dual-layer collaborative recognition architecture with a main channel Transformer model and an auxiliary channel CTC-Attention hybrid architecture to obtain text with multi-dimensional feature annotations.

[0009] Based on text with multi-dimensional feature annotations, it uses a bidirectional recurrent neural network to perform scene classification and dependency syntax analysis, and combines it with a dynamic knowledge graph to retrieve related entities in real time to generate structured intent data.

[0010] The structured intent data is analyzed in three dimensions through the explainability engine to generate a multi-dimensional traceability report.

[0011] Based on multi-dimensional traceability reports, the intelligent decision fusion engine is triggered by confidence thresholds to generate structured enhanced interaction data;

[0012] Based on structured enhanced interaction data, multimodal interaction is carried out through dynamic traceability visual interface rendering and tactile coding feedback of smart wearable devices, and a complete interaction log is generated.

[0013] As a preferred solution of the voice interaction method based on artificial intelligence technology described in the present invention, wherein: the deep neural network noise reduction model is used to eliminate environmental noise, and the specific steps are as follows:

[0014] A deep neural network denoising model is constructed based on a hybrid framework of convolutional neural networks and recurrent neural networks.

[0015] The convolutional neural network layer extracts the spectral features of the speech signal after beamforming, and the recurrent neural network layer captures the temporal features of the speech signal to distinguish it from noise;

[0016] A deep neural network denoising model is trained using paired data of noisy and clean speech, and the difference between the denoised features and the clean speech is calculated using the mean square error loss function.

[0017] The Adam optimizer is used to back-propagate and update parameters. After optimization on the validation set, the trained deep neural network denoising model is deployed to estimate and suppress the noise components. The output layer outputs the denoised speech features, which are then encoded to form a voice data packet.

[0018] As a preferred solution of the voice interaction method based on artificial intelligence technology described in the present invention, the dual-layer collaborative recognition using the main channel Transformer model and the auxiliary channel CTC-Attention hybrid architecture is specifically performed as follows:

[0019] After preprocessing, the speech processing packet is input into the Transformer model of the main channel and the CTC-Attention model of the auxiliary channel for feature encoding;

[0020] The Transformer encoding process generates a context representation through multi-head self-attention, then performs residual connection and layer normalization, and then performs residual connection and layer normalization twice after nonlinear transformation through a feedforward neural network, and finally outputs a coded feature sequence;

[0021] The attention mechanism dynamically calculates the interaction weight between each character in the target text and the encoded feature sequence and generates a context vector. The time step features of the CTC-Attention model are concatenated with the attention context vector to generate an optimized feature sequence.

[0022] The main channel and auxiliary channel are reconstructed into a global semantic strategy layer and a local alignment execution layer through functional decoupling. The global semantic strategy layer focuses on cross-frame semantic associations, while the local alignment execution layer strengthens the alignment confidence at the time step granularity.

[0023] A gated residual fusion strategy is adopted to dynamically fuse global semantic features with local alignment features to generate multi-dimensional feature-annotated text.

[0024] As a preferred solution of the voice interaction method based on artificial intelligence technology described in the present invention, wherein: the real-time retrieval of related entities in combination with the dynamic knowledge graph to generate structured intent data is carried out in the following specific steps:

[0025] Extract scene classification results and dependency information from text with multi-dimensional feature annotations, dynamically combine them into query conditions, and retrieve related entities in the knowledge graph in real time;

[0026] Relevant entities are filtered through entity matching and similarity calculation, and integrated into structured intent data in JSON format.

[0027] The three-dimensional analysis is performed through the explainability engine to generate a multi-dimensional traceability report. The specific steps are:

[0028] The structured intent data is input into the explainability engine. Through hierarchical feature decoupling, the structured intent data is split into three independent encoder channels: intent dimension encoder channel, entity association dimension encoder channel, and time series trajectory dimension encoder channel. The split data are processed separately.

[0029] The semantic confidence vector, entity association strength vector and temporal feature vector output by the three-channel independent encoder are enhanced by residual connection and input into the Gram-Schmidt orthogonalization process to eliminate the linear correlation between dimensions and generate a mutually independent orthogonal basis vector group. A multi-dimensional traceability report is generated based on the orthogonal basis vector group.

[0030] As a preferred solution of the voice interaction method based on artificial intelligence technology described in the present invention, wherein: the intelligent decision fusion engine is triggered by the confidence threshold to generate structured enhanced interaction data, the specific steps are:

[0031] Set semantic dimension thresholds, context dimension thresholds, and intent dimension thresholds, compare the confidence levels of each dimension, and dynamically trigger the intelligent decision fusion engine.

[0032] The dependency annotations and associated entity annotations in the semantic dimension are fused to enhance the semantic representation vector. The scene classification results and context labels in the context dimension are context-integrated to enhance the context representation vector. The intent labels and intent verification results in the intent dimension are combined through a weighted fusion operation to optimize the intent representation vector.

[0033] The fused semantic dimension, context dimension and intention dimension are integrated through data aggregation, and JSON standard encoding is used to generate structured enhanced interaction data.

[0034] As a preferred solution of the voice interaction method based on artificial intelligence technology described in the present invention, wherein: the multimodal interaction is performed through dynamic traceability visual interface rendering and tactile coding feedback of smart wearable devices, and a complete interaction log is generated. The specific steps are:

[0035] The dynamic traceability visualization interface renders a three-layer interactive view based on enhanced interaction data, and the smart wearable device tactile encoding engine converts entity association strength and intent confidence into dual-channel tactile signals;

[0036] Through timestamp synchronization, tactile pulses are linked to visual animation keyframes, and pressure sensor data reversely regulates the interface mechanical parameters in real time to form a closed-loop interaction.

[0037] The timeline-aligned voice data stream, visual operation path, and tactile parameter triples are encoded according to the W3CActivityStreams 2.0 specification and verified with SHA-256 to generate a complete interaction log that can be restored across devices.

[0038] In a second aspect, the present invention provides a voice interaction system based on artificial intelligence technology, comprising: an acquisition module, a labeling module, a parsing module, an analysis module, a fusion module and an interaction module.

[0039] The acquisition module captures user voice data, synchronizes the sound source focusing algorithm to lock the target direction, and uses a deep neural network noise reduction model to eliminate environmental noise and output voice data packets;

[0040] The annotation module uses a dual-layer collaborative recognition architecture consisting of a main channel Transformer model and an auxiliary channel CTC-Attention hybrid architecture to obtain text with multi-dimensional feature annotations.

[0041] The parsing module, based on text with multi-dimensional feature annotations, performs scene classification and dependency syntax analysis through a bidirectional recurrent neural network, and combines it with a dynamic knowledge graph to retrieve related entities in real time to generate structured intent data;

[0042] The analysis module uses the explainability engine to perform three-dimensional analysis on structured intent data and generate a multi-dimensional traceability report.

[0043] The fusion module, based on multi-dimensional traceability reports, triggers the intelligent decision fusion engine through confidence thresholds to generate structured enhanced interaction data;

[0044] The interaction module, based on structured enhanced interaction data, conducts multimodal interaction through dynamic traceability visual interface rendering and tactile coding feedback of smart wearable devices, and generates a complete interaction log.

[0045] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the voice interaction method based on artificial intelligence technology as described in the first aspect of the present invention is implemented.

[0046] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the voice interaction method based on artificial intelligence technology as described in the first aspect of the present invention.

[0047] The beneficial effects of the present invention are: eliminating environmental noise through a deep neural network noise reduction model, adopting a two-layer collaborative recognition of the main channel Transformer model and the auxiliary channel CTC-Attention hybrid architecture, achieving high-precision recognition and multi-dimensional feature labeling of speech signals, performing scene classification and dependency syntax analysis through a bidirectional recurrent neural network, and combining with a dynamic knowledge graph to retrieve related entities in real time and generate structured intent data, thereby accurately understanding user intentions and providing personalized interactive services. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0049] Figure 1 The figure is a flow chart of a voice interaction method based on artificial intelligence technology.

[0050] Figure 2 Flowchart of the deep neural network denoising model.

[0051] Figure 3 Flowchart of collaborative identification between main channel and auxiliary channel.

[0052] Figure 4 This is the core flow chart of multimodal interaction. DETAILED DESCRIPTION

[0053] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0054] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0055] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0056] Reference Figures 1 to 4 , is an embodiment of the present invention, which provides a voice interaction method based on artificial intelligence technology, including the following steps:

[0057] S1: Capture user voice data, synchronize the sound source focusing algorithm to lock the target direction, and use the deep neural network noise reduction model to eliminate environmental noise, and output the voice data packet;

[0058] The user's voice data is collected through a multi-microphone array. The microphone array is arranged in a ring, which can evenly cover all directions and is suitable for all-round sound collection. The signal collected by each microphone is stored in a digital audio format.

[0059] Beamforming technology is used to pre-process the voice signals collected by multiple microphones. By calculating the time delay and phase difference of each microphone signal, the voice signal is focused to a specific direction and the direction of the target sound source is locked. In a multi-microphone array, based on the speed of sound waves in the air, the time difference between any two microphones receiving the signal is calculated. The time difference reflects the difference in the path length of the sound wave from the sound source to the two microphones divided by the speed of sound. Based on the calculated arrival time difference and the position information of the microphone in the array, the weighting coefficient of each microphone signal is determined. The weighting coefficient is dynamically adjusted according to the position of the microphone and the direction of the target sound source, so that the signal is enhanced in the direction of the target sound source and attenuated in other directions. The coefficients perform weighted processing on the digital audio signal collected by each microphone, and adjust the amplitude of each microphone signal so that the signal in the direction of the target sound source can be highlighted, and the signals in other directions are suppressed. The weighted signal retains the characteristics of the target sound source while reducing the influence of background noise and other interference sources. All microphone signals after weighted processing are superimposed. Since the weighted processing has adjusted the amplitude and phase of each signal, the signals from the direction of the target sound source can enhance each other when superimposed, forming a beam pointing to the target sound source. The beam pointing to the target sound source is directional and can be focused on the target sound source, thereby locking the direction of the target sound source. The directionality of the beam can determine the position of the target sound source in space;

[0060] Using a hybrid architecture of convolutional neural networks and recurrent neural networks, the beamforming speech signal is input into a deep neural network noise reduction model. The convolutional neural network is used to extract features from the input speech signal. The convolutional layer in the convolutional neural network extracts local features using a sliding window approach, and the pooling layer is used to reduce the dimension of the features while retaining important information. The convolutional and pooling layers in the convolutional neural network automatically learn the spectral features of the speech signal. The spectral features include information about the frequency distribution and energy distribution of the speech signal and can identify the spectral differences between the speech signal and the noise signal.

[0061] The features extracted by the convolutional neural network are input into the recurrent neural network. The recurrent neural network can capture the temporal continuity and changing patterns of the speech signal. The recurrent neural network can remember the information at the previous moment and use it for processing at the current moment, better understanding the dynamic changes of the speech signal and further distinguishing between noise and speech signals.

[0062] A large number of noisy speech samples and corresponding clean speech samples are collected as training data sets. Noisy speech samples refer to speech signals containing background noise recorded in a real environment, and clean speech samples refer to pure speech signals recorded in a noise-free environment. The long speech signal is divided into shorter frames, and the spectral features of each frame of speech signal are extracted. By adding different types of noise and adjusting the signal-to-noise ratio, the training data set is further expanded to construct a deep neural network denoising model with a hybrid architecture based on convolutional neural networks and recurrent neural networks. The input layer receives the preprocessed noisy speech features, the convolutional neural network layer is used to extract the spectral features of the speech signal, and the recurrent neural network layer is used to capture the temporal features of the speech signal. The output layer outputs the denoised speech features. A loss function is defined to measure the difference between the denoised speech features and the clean speech features output by the model. Commonly used loss functions include mean square error loss. The square error B between the model output and the target clean speech features is calculated as follows:

[0063]

[0064] Where N is the number of samples, y i is the clean speech feature, is the noise reduction speech feature output by the model, and i is the index of the i-th sample in the training dataset;

[0065] Use the training data set to train the deep neural network denoising model, input the noisy speech features into the deep neural network denoising model, output the denoised speech features through spectral features and time series features, use the loss function to calculate the error between the model output and the clean speech features, and update the parameters of the deep neural network denoising model through the back propagation algorithm based on the calculated loss to minimize the loss function. Select the Adam optimization algorithm to adjust the parameters of the model. The Adam optimization algorithm gradually adjusts the weights of the model according to the gradient of the loss function to achieve the optimal denoising effect. During the training process, use the validation data set to verify the deep neural network denoising model and evaluate the performance. According to the validation results, adjust the model's hyperparameters, optimize the model's performance, and save it. The validation data set is a part of the data divided from the training data set, which is used to simulate the performance of the model on unknown data and output the trained deep neural network denoising model.

[0066] The trained deep neural network noise reduction model is used to estimate and remove the noise components in the speech signal. A large number of noisy speech samples and corresponding clean speech samples are used to optimize the parameters of the deep neural network noise reduction model by minimizing the noise estimation error, so that the deep neural network noise reduction model can accurately identify and remove the noise components. In practical applications, the deep neural network noise reduction model outputs a noise-reduced speech signal based on the input noisy speech signal. The speech signal processed by the deep neural network noise reduction model is encoded and packaged to form a voice data packet. The voice data packet contains metadata information such as the sampling rate and quantization bit number of the speech signal and is stored in a digital audio format.

[0067] S2: Uses a dual-layer collaborative recognition architecture with a main channel Transformer model and an auxiliary channel CTC-Attention hybrid architecture to obtain text with multi-dimensional feature annotations;

[0068] Preprocess the voice data packets to adapt them to the input requirements of the main channel Transformer model. Segment the voice data packets into a fixed-length frame sequence and perform Fourier transform on each frame signal to obtain spectral features. Filter the spectral features using a Mel filter bank and calculate the energy value. Perform discrete cosine transform on the energy value to extract Mel-frequency cepstral coefficient features. Normalize the extracted spectral features and Mel-frequency cepstral coefficient features to obtain Mel-frequency cepstral feature values. Scale the Mel-frequency cepstral feature values to a fixed range, such as between 0 and 1, and input the normalized feature values into the main channel Transformer model. Embed the normalized feature values of each frame through a learnable linear transformation, elevating the low-dimensional normalized features to a high-dimensional space to generate an embedded feature vector for each frame.

[0069] During the model training phase, the preprocessed speech feature sequence is trained end-to-end jointly. The main channel Transformer model encodes the embedded feature vectors through a multi-head self-attention mechanism. The entire embedded vector sequence is batch-projected into a query vector matrix, a key vector matrix, and a value vector matrix through three sets of independently learnable weight matrices. The attention score is calculated by the dot product of the query vector matrix and the key vector matrix. After scaling, it is normalized into attention weights using the Softmax function and then weighted summed with the value vector matrix to generate an intermediate context representation. After processing by the multi-head self-attention mechanism, the Transformer model adds the residuals of the input and output of the multi-head self-attention layer and applies layer normalization to the addition result to adjust the mean and variance of the features. The Transformer model performs a nonlinear transformation on the normalized features through a feedforward neural network. The feedforward neural network is processed in sequence through two fully connected layers. The ReLU activation function is applied in the intermediate layer. The input and output of the feedforward neural network are again residually added. The layer normalization operation is repeated on the addition result, and the encoded feature sequence is finally output.

[0070] At the same time, the encoded feature sequence is input into the CTC-Attention model of the auxiliary channel. The shape and length of the encoded feature sequence are adjusted to match the input requirements of the CTC-Attention model. The encoded feature sequence is forward propagated through the CTC-Attention model to generate the output probability distribution of each time step, forming an output probability matrix. Each element of the matrix represents the predicted probability of the CTC-Attention model for a specific character. The target text is converted into a character sequence. All possible alignment paths are enumerated through the dynamic programming algorithm, and the probability P(π|X) of each path is calculated. The formula is:

[0071]

[0072] Among them, X represents the encoded feature sequence input to the CTC-Attention model, T represents the length of the input sequence, and π t represents the character corresponding to the alignment path π at time step t, Indicates that at time step t, the CTC-Attention model is on the character π t The predicted probability of

[0073] Calculate the total probability value P(Y|X) of each alignment path to get the alignment error. The formula is:

[0074] P(Y|X)=∑ π∈Π P(π|X);

[0075] A = 1-P(Y|X);

[0076] Where Y represents the character sequence of the target text, π represents the set of all possible alignment paths, and A represents the alignment error;

[0077] The total probability value is converted into a CTC loss value by taking the negative log-likelihood. The smaller the CTC loss value, the higher the degree of match between the model output and the target text. The formula is:

[0078] CTCL = -log P(Y|X);

[0079] Among them, CTCL represents the CTC loss value;

[0080] Based on the CTC-Attention model, the attention mechanism is introduced to further optimize feature alignment. The attention weight between each character in the target text and the feature vector of each time step in the input encoded feature sequence is calculated and normalized to represent the contribution of the input encoded feature sequence to the target text. The calculated attention weight is used to perform weighted summation on the input encoded feature sequence to generate a context vector.

[0081] After completing the CTC loss calculation and attention mechanism optimization, the feature vector of each time step is merged with the corresponding context vector by splicing to generate an optimized feature vector. The integrated feature vectors are arranged in the order of time steps to generate an optimized feature sequence.

[0082] The main channel Transformer model and the auxiliary channel CTC-Attention model are reconstructed into a global semantic strategy layer and a local alignment execution layer through functional decoupling. The global semantic strategy layer focuses on cross-frame semantic associations, while the local alignment execution layer strengthens the alignment confidence at the time step granularity.

[0083] The global semantic strategy layer processes the input encoding feature sequence through a multi-head self-attention mechanism and a feedforward neural network to generate global semantic features that contain long-distance dependencies between phonemes. The local alignment execution layer generates local phoneme boundary features based on the optimized feature sequence of CTC path probability and attention weight optimization alignment. The two-layer confidence is quantified by calculating the self-attention weight entropy of the global semantic strategy layer and the CTC loss reverse mapping value of the local alignment execution layer. A gated residual fusion strategy is adopted to dynamically fuse the global semantic features with the local alignment features. Finally, the fused features are multi-dimensionally annotated with phonemes, semantics, and context through a multi-task joint framework to generate multi-dimensional feature-annotated text with cross-granularity alignment capabilities.

[0084] S3, based on text with multi-dimensional feature annotations, performs scene classification and dependency syntax analysis through a bidirectional recurrent neural network, and combines it with a dynamic knowledge graph to retrieve related entities in real time to generate structured intent data;

[0085] Extract the phoneme features, semantic features, and context features of each time step from the text with multi-dimensional feature annotations, and organize them into a multi-dimensional feature vector sequence through feature fusion. Input the extracted multi-dimensional feature vector sequence into the forward recurrent neural network of the bidirectional recurrent network, process the context information from left to right, initialize the hidden state of the forward recurrent neural network, usually set to a zero vector, and then process the multi-dimensional feature vector of each time step from front to back, and update the hidden state of the forward recurrent neural network. The hidden state of each time step depends on the input of the current time step and the hidden state of the previous time step. After processing by the forward recurrent neural network, a forward context information feature sequence is generated. The multi-dimensional feature vector of each time step contains the context information from the beginning of the sequence to the current time step.

[0086] The extracted multi-dimensional feature vector sequence is input into the backward recurrent neural network of the bidirectional recurrent network, and the context information from right to left is processed. The hidden state of the backward recurrent neural network is initialized, usually set to a zero vector. Then the multi-dimensional feature vector of each time step is processed from back to front, and the hidden state of the backward recurrent neural network is updated. The hidden state of each time step depends on the input of the current time step and the hidden state of the next time step. After processing by the backward recurrent neural network, a feature sequence of backward context information is generated. The feature vector of each time step contains the context information from the current time step to the end of the sequence.

[0087] For each time step, the multidimensional feature vectors generated by the forward recurrent neural network and the backward recurrent neural network are concatenated to form a comprehensive multidimensional feature vector, and arranged in chronological order to form a bidirectional feature sequence;

[0088] Receive the bidirectional feature sequence processed by the bidirectional recurrent network, input the bidirectional feature vector of each time step into the fully connected layer, the output dimension of the fully connected layer is consistent with the number of scene categories, and the bidirectional feature vector is mapped to the space of scene categories through the fully connected layer to generate the scene category score corresponding to each time step. Apply the Softmax function to the scene category score output by the fully connected layer, and convert the scene category score into a probability value. The probability value represents the predicted probability of each scene category at the corresponding time step. Arrange the predicted probabilities of all time steps in time step order to form a probability matrix, where each row corresponds to a time step and each column corresponds to a scene category. Each element in the probability matrix is the predicted probability of the corresponding scene category at the corresponding time step. Analyze the generated prediction probability matrix, calculate the average probability of each scene category along the time dimension for the prediction probability matrix of all time steps, and select the scene category with the highest average probability as the scene classification result.

[0089] Preprocess the text with multi-dimensional feature annotations, split the preprocessed text with multi-dimensional feature annotations into separate words, perform part-of-speech tagging on each word, determine the grammatical role of each word in the sentence, and convert it into a vector form suitable for dependency syntax analysis model processing. Load the pre-trained dependency syntax analysis model based on rule-based methods and machine learning models, collect large-scale corpus with dependency relationship annotations through data preparation and annotation, and perform word segmentation, part-of-speech tagging and dependency relationship triple tagging on the corpus, and load and initialize the dependency syntax analysis model. The rule method requires defining a grammar rule library and Dependency matching is achieved through hard coding, while the machine learning method loads the pre-trained dependency parsing model. The configuration file and weights are loaded through the pre-training method, and the pre-trained dependency parsing model is fine-tuned using domain data to adapt to the task. Then, the feature input and parsing phase begins. The vectorized text is input into the dependency parsing model, and the classifier predicts the head word index and dependency type. The prediction results are post-processed in combination with the rule base to correct conflicting dependency edges. Finally, the output and integration are completed, the prediction results for each word are output, and the prediction results are back-annotated to the original text to generate enhanced dependency annotation text.

[0090] Based on the prediction results, a dependency list containing subordinate word index, head word index and dependency relationship type is extracted, and the dependencies are sorted according to the subordinate word index to ensure that their processing order is consistent with the order of words in the sentence. In the process of building the dependency tree, the root node of the sentence is identified as the starting point of the dependency tree. The root node has no head word and represents the main structural center of the sentence. Based on the root node, the dependency tree structure is initialized to store nodes and edges. The root node is added to the initialized dependency tree structure. The root node becomes the first node of the dependency tree, and the previously sorted dependencies are added. For each dependent word in the storage relationship list, find the corresponding head word node in the dependency tree according to the head word index, add a node for the dependent word in the dependency tree, and create an edge from the head word node to the dependent word node. Each node represents a word, and the edge represents the dependency relationship. The entire dependency tree is gradually constructed. After adding each dependent word node, the structure of the dependency tree is updated so that the hierarchical relationship of the dependency tree can accurately reflect the grammatical and semantic levels of the sentence. Based on the text with multi-dimensional feature annotations, the annotation information of the dependency relationship is added as the result of dependency syntactic analysis;

[0091] Extract the scene category to which the text belongs from the scene classification results, extract the semantic entity identifiers and relationship types from the dependency syntax analysis results, combine the scene category and semantic information into query conditions, initialize the dynamic knowledge graph query based on the query conditions, and in the dynamic knowledge graph, query related entities and relationships based on the scene category and dependency relationship information, and generate query results;

[0092] Extract entities related to the text content from the query results, match the extracted entities with the text content, and associate the entities with words in the text. Based on dependency information, further filter out entities that are highly relevant to the text content. Use similarity calculations to evaluate the relevance of each entity to determine the degree of match with the text content. Organize the identifier, type, and related attributes of each filtered entity into entity information. Combine all organized entity information to generate associated entity annotations.

[0093] Integrate the scene classification results, dependency annotation text and associated entity annotations into a unified data structure, parse the scene classification results, dependency annotation text and associated entity annotations in the unified data structure, extract the key information of scene categories, main dependencies and key entities, and organize the extracted key information according to the JSON structured data format to form structured intent data.

[0094] S4. Perform three-dimensional analysis on the structured intent data through the explainability engine to generate a multi-dimensional traceability report.

[0095] The structured intent data is fed into the explainability engine to perform hierarchical feature decoupling. Three independent encoders are used to extract intent dimension features, entity association dimension features, and temporal trajectory dimension features respectively.

[0096] The intent dimension encoder channel loads a pre-trained self-attention mechanism layer, performs multi-head attention calculation on the input intent label sequence, and generates a semantic confidence vector. The number of attention heads strictly corresponds to the number of intent label categories, and the attention weight matrix is normalized using the Softmax function.

[0097] The entity association dimension encoder channel activates the graph convolutional network layer, which uses the knowledge graph entity ID as the node and the dependency path as the edge to build the graph structure, perform two-hop neighborhood information aggregation, and use mean pooling to aggregate neighbor information for node feature update, and output the entity association strength vector;

[0098] The encoder channel of the temporal trajectory dimension deploys a bidirectional long short-term memory network layer, which inputs the temporal trajectory dimension event continuity indicator sequence in timestamp order. The forward and reverse hidden states are concatenated to generate a temporal feature vector. The output vector of each encoder channel is enhanced by residual connection and then input into the Gram-Schmidt orthogonalization process to eliminate the linear correlation between dimensions and generate a group of independent orthogonal basis vectors.

[0099] The orthogonalized feature vectors are subjected to dynamic weight allocation. Monte Carlo sampling is used to simulate feature-missing scenarios. The difference in the impact of each dimension on the decision outcome is calculated. After normalization, the weights of the intent dimension, entity association dimension, and time series trajectory dimension are generated. When the weight coefficient is lower than the dynamic average threshold, the feature is marked as low contribution, and the orthogonalized copy is retained. The high-weight feature and the orthogonalized copy are spliced together to generate a decoupled feature vector group.

[0100] The decoupled feature vector group is input into the two-layer rule engine, and the rule base is constructed through the C4.5 algorithm and DBSCAN clustering. The matching tasks are sorted by feature weights and the temporal continuity, entity relevance and meaning are performed. Figure 1 The three-stage verification of consistency ultimately generates a structured decision path. At the same time, anomalies are fed back to the feature decoupling layer to dynamically adjust weights to achieve closed-loop optimization.

[0101] After the verified structured decision path enters the dynamic traceability stage, a traceability path is constructed. Intent dimension traceability uses the unique identifier of the trigger node in the rule base path and the confidence change trajectory to generate a node trigger chain in chronological order.

[0102] Entity dimension traceability extracts the diffusion paths of core entities and edge entities in the knowledge graph, marks the attenuation trend of the association strength during the diffusion process, and forms an entity relationship strength change curve;

[0103] Time series dimension traceability builds an event chain causal relationship diagram and marks the timestamp offset of key event nodes;

[0104] Multimodal reports are synthesized based on the traceability path results. Text reports generate natural language descriptions according to the decision conclusion-rule basis-abnormal warning structure. Visual maps use a force-directed graph layout to display the relationship between entities and events. Node size represents feature weight, and edge transparency reflects the strength of temporal association. A timeline heat map is superimposed to represent confidence changes with color gradients. Interactive logs fully record the application interface call sequence and intermediate state snapshots during the decision-making process, and support rolling back to any historical moment along the timeline to view data status.

[0105] The node trajectory, entity attenuation curve and timestamp offset output in the traceability path construction phase are directly used as the input data of the multimodal report, and the visual map and interactive log are associated through a unified timeline. Clicking on the map node can locate the corresponding operation record in the log, and finally generate a multi-dimensional traceability report. Among them, the multi-dimensional traceability report includes three parts: structured text report, visual force-directed map and interactive log. The three are associated through a unified timeline to achieve collaborative analysis and backtracking of text description, map display and operation records.

[0106] S5. Based on the multi-dimensional traceability report, the confidence threshold is used to trigger the intelligent decision fusion engine to generate structured enhanced interaction data;

[0107] A sliding window algorithm is used to process the distribution data of historical semantic verification confidence and semantic consistency confidence. The sliding window algorithm traverses historical data in fixed time windows, calculates the mean of the semantic verification confidence and semantic consistency confidence within each window, and subtracts one standard deviation from the mean as a dynamic benchmark value to generate the semantic dimension threshold.

[0108] Based on the historical scene classification confidence data output by semantic dimension threshold initialization, kernel density estimation is applied to fit the probability density distribution of this data. Kernel density estimation smoothes the data using a Gaussian kernel function to generate a probability density curve. The inflection point of the curve where the probability density decreases is identified to generate the context dimension threshold.

[0109] The momentum gradient descent algorithm is used to process the time series data of the product of intent completeness confidence and accuracy confidence. The momentum gradient descent algorithm iteratively calculates the gradient of the product value, tracks its fluctuation boundary, and locks the lower limit of the boundary to generate the intent dimension threshold.

[0110] Read the semantic confidence, context confidence, and intent confidence from the multi-dimensional traceability report, and compare the semantic confidence with the semantic dimension threshold, the context confidence with the context dimension threshold, and the intent confidence with the intent dimension threshold. When the confidence value of any dimension is less than the corresponding threshold, a signal is triggered to activate the intelligent decision fusion engine.

[0111] Perform dependency syntactic analysis on the original text, automatically annotate it based on the deep learning dependency syntactic analysis method, generate a set of grammatical dependencies as dependency annotations, extract entity attributes related to the dependency annotations from the knowledge graph, generate a structured entity attribute vector, and splice the dependency annotations with the associated entity annotations using the feature dimension splicing method to obtain an enhanced semantic representation vector;

[0112] The scene classification results are mapped into a sequence of integer indices through ordered label encoding, which is then converted into a category encoding vector. The context labels are concatenated into the original context vector according to the feature dimension. The scene category encoding vector and the original context vector are weighted summed to obtain a context enhancement matrix.

[0113] The intent label and intent verification result are spliced into the original intent vector through the feature dimension splicing method, and the intent label and verification result are combined through weighted fusion operation to obtain the optimized intent representation vector;

[0114] The semantic enhancement vector, context enhancement matrix, and intent enhancement vector are decomposed into a unified field structure according to feature dimensions using a standardized feature dimension decomposition method. The entity attributes of the semantic enhancement vector and the environmental parameters of the context enhancement matrix are stored independently. The intent enhancement vector is aligned with the scene classification probability as the confidence field.

[0115] The aligned field structure is spliced through the feature dimension key-value pair mapping method, and the triples of the semantic enhancement vector are converted into key-value pairs; the numerical fields of the context enhancement matrix are stored independently without merging the semantic entity attributes; the intent enhancement vector is inserted as an independent field, and the spliced fields are integrated through the nested JSON construction method to generate a nested JSON structure. The Industrial Internet of Things data specification is adopted to generate structured enhanced interactive data.

[0116] S6. Based on structured enhanced interaction data, multimodal interaction is performed through dynamic traceability visual interface rendering and tactile coding feedback of smart wearable devices, and a complete interaction log is generated;

[0117] The enhanced interaction data is input into the multimodal distributor, and data field decoupling is performed. The semantic enhancement field, timestamp correction mark and intent metadata field in the enhanced interaction data are received. The semantic enhancement field generates a priority queue through the entity association strength weight sorting algorithm. Each element in the queue consists of a binding pair of the dependency path and the knowledge graph entity ID. The priority weight is determined by the inverse of the entity association strength value and the square of the dependency path attention weight. The timestamp correction mark uses a sliding window mechanism to align the time axis of the original voice data. The window size is dynamically adjusted according to the average response delay of the most recent interactions. The timestamps in the window are reallocated by linear interpolation to eliminate the timing deviation between multimodal data. The intent metadata field is parsed by the rule engine to generate an intent state transition matrix. The rows and columns of the intent state transition matrix correspond to the current intent label and the next scene classification label, respectively. The matrix element value is the statistical value of the number of rule triggering times, which is mapped to the 0-1 interval after normalization.

[0118] A three-layer visualization rendering engine is constructed. The entity relationship layer inputs the priority queue into the force-directed graph generator. The repulsion coefficient between nodes is dynamically calculated based on the inverse of the entity association strength value. The edge weight is determined by the normalized result of the dependency path attention weight. When the change in entity association strength exceeds the historical standard deviation, the node position interpolation animation is triggered. The animation duration is inversely proportional to the change amplitude, making the visual transition smooth. The time series trajectory layer inputs the timestamp sequence after the time axis sliding window is aligned and corrected into the heat map generator. The Gaussian kernel density estimation algorithm is used to estimate the context coverage of each time slot. The heat value density calculation formula is:

[0119] D = ln(C + 1) × k;

[0120] Where D represents the thermal value density, C represents the context coverage, and k represents the dynamic scaling factor;

[0121] After receiving the thermal density, the Canvas2D rendering engine is used to draw the timeline baseline. The transparency decays linearly with the distance from the current time point. The intent projection layer maps the intent state transfer matrix to three-dimensional space using a dimensionality reduction algorithm. The gravitational constraint mechanism calculates the spatial distribution based on the intent transfer probability and the surface area of the sphere. When the user drags the sphere, the neighboring spheres are rearranged based on the dynamic balance of gravity. The sphere radius is adjusted in real time and maintains a positive correlation with the confidence level of the intent dimension. The sphere corresponding to high-confidence intent appears larger.

[0122] Extract intent dimension confidence and context coverage from enhanced interaction data. JSON data standardization is used to eliminate dimensional differences. The parsed data is stored in a dual-channel buffer, with intent data allocated to a quick access area and context data to a large-capacity storage area. Tactile pattern matching uses an improved similarity algorithm to match waveform templates in a pre-set tactile library. A dynamic adjustment mechanism automatically adjusts the matching weight based on context coverage and outputs the top three candidate patterns.

[0123] The base frequency automatically adapts to the preset interval based on the confidence level of the intent dimension. The amplitude strength is dynamically adjusted through an exponential function combined with the scenario tag. The pulse duration is set based on the sliding window average of the recent interactive response time. The generated parameter group is converted into a standard instruction format after verification. The standard instruction format after the verification of the parameter group is fragmented and encapsulated using a fragmentation encapsulation strategy. Data packets are transmitted according to the intent priority, and high-priority instructions can preempt the transmission channel.

[0124] After receiving the parsed command parameters, the linear resonant motor loads the resonant frequency matching table during the initialization phase to establish the motor coil impedance characteristic curve. During the amplitude control phase, the pulse width modulation duty cycle is adjusted to output the target vibration intensity. The timing controller synchronizes and coordinates the superposition of multi-channel tactile effects, and uses a hardware-accelerated interpolation algorithm to phase-align and smooth the transition of waveforms from different channels. During this process, the pressure sensor array collects deformation distribution data of the contact surface in real time, and feeds the pressure gradient information back to the calibration loop for dynamic parameter compensation. The calibrated data is then synchronously updated to the force feedback rendering engine of the dynamic traceability visualization interface, forming a closed-loop control that integrates tactile and visual perception.

[0125] A collaborative control mechanism for tactile feedback and a dynamic traceability visualization interface is established. Timestamp alignment technology is used to synchronize the triggering of the tactile pulse start edge with the keyframes of the visualization particle animation. The contact surface deformation data collected by the pressure sensor array is filtered and then fed back into the force-directed graph rendering engine of the visualization interface. The edge weight parameters of the knowledge graph entity nodes are dynamically adjusted, and calibration instruction data is used to achieve multi-system collaboration through shared memory.

[0126] The embedded coprocessor monitors coil temperature, current peak, and skin impedance fluctuations in real time. In case of anomalies, it sequentially executes amplitude attenuation, switches to alternate waveforms, and cuts off the power supply. The monitoring threshold is dynamically calculated based on historical impedance data, and the safety status indicator on the visual interface is simultaneously updated, forming a closed-loop control loop between device operating parameters and user protection.

[0127] The event timestamps of the dynamic traceability visualization interface and the tactile feedback timestamps of the smart wearable device are synchronized and calibrated through the improved NTP protocol to generate a unified time base axis. The voice data flow is divided into independent interaction units through endpoint detection. Each unit is bound to the visualization operation path coordinates and tactile feedback parameters in the corresponding time period to form a time-aligned multimodal data set. Among them, the tactile feedback parameters include frequency, amplitude and duration. The voice unit is converted into UTF-8 text after recognition, the visualization operation record is extracted as a path coordinate sequence, and the tactile feedback parameters are parsed into structured triple data. The structured triple data is indexed and associated through a shared timestamp to construct heterogeneous data blocks with time tags. Based on the W3CActivityStreams2.0 specification, the voice interaction entity, visualization operation entity, and tactile feedback entity are defined and connected through the "trigger" and "include" semantic relationships. Data fusion Timely detection of timestamp overlap conflicts, priority is given to retaining the time boundary of voice data, the conflict resolution algorithm dynamically adjusts the entity association relationship according to the interaction type weight, and encodes the fused data according to the JSON specification. The core fields include semantic context, device identification, interaction data entity and validity evaluation indicators. After adding the SHA-256 checksum, hierarchical storage is implemented, recent logs are stored in the memory database, and historical logs are written to distributed storage after compression. Access policy metadata is attached during encrypted storage, and access requires verification of dual credentials of device fingerprint and cloud token. Abnormal requests return desensitized summaries processed with differential privacy. A standardized query interface is generated through OpenAPI3.0, which supports natural language time expression retrieval. The query results return a replayable data package, where the replayable data package contains the original voice waveform, visual operation animation timeline, and tactile parameter instruction set, supporting the restoration of the complete interaction process across devices.

[0128] This embodiment also provides a voice interaction system based on artificial intelligence technology, including: an acquisition module, a labeling module, a parsing module, an analysis module, a fusion module and an interaction module.

[0129] The acquisition module captures user voice data, synchronizes the sound source focusing algorithm to lock the target direction, and uses a deep neural network noise reduction model to eliminate environmental noise and output voice data packets;

[0130] The annotation module uses a dual-layer collaborative recognition architecture consisting of a main channel Transformer model and an auxiliary channel CTC-Attention hybrid architecture to obtain text with multi-dimensional feature annotations.

[0131] The parsing module, based on text with multi-dimensional feature annotations, performs scene classification and dependency syntax analysis through a bidirectional recurrent neural network, and combines it with a dynamic knowledge graph to retrieve related entities in real time to generate structured intent data;

[0132] The analysis module uses the explainability engine to perform three-dimensional analysis on structured intent data and generate a multi-dimensional traceability report.

[0133] The fusion module, based on multi-dimensional traceability reports, triggers the intelligent decision fusion engine through confidence thresholds to generate structured enhanced interaction data;

[0134] The interaction module, based on structured enhanced interaction data, conducts multimodal interaction through dynamic traceability visual interface rendering and tactile coding feedback of smart wearable devices, and generates a complete interaction log.

[0135] This embodiment also provides a computer device suitable for the voice interaction method based on artificial intelligence technology, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the voice interaction method based on artificial intelligence technology proposed in the above embodiment.

[0136] The computer device may be a terminal, comprising a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner may be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a button, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse.

[0137] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the voice interaction method based on artificial intelligence technology proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, disk or optical disk.

[0138] In summary, the present invention achieves high-precision recognition and multi-dimensional feature labeling of speech signals through: eliminating environmental noise with a deep neural network denoising model, adopting a dual-layer collaborative recognition of a main channel Transformer model and an auxiliary channel CTC-Attention hybrid architecture, performing scene classification and dependency syntax analysis through a bidirectional recurrent neural network, and combining with a dynamic knowledge graph to retrieve related entities in real time and generate structured intent data, thereby accurately understanding user intent and providing personalized interactive services.

[0139] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A voice interaction method based on artificial intelligence technology, characterized by: include: Capture user voice data, synchronize the sound source focusing algorithm to lock the target direction, and use the deep neural network noise reduction model to eliminate environmental noise and output voice data packets; It uses a dual-layer collaborative recognition architecture with a main channel Transformer model and an auxiliary channel CTC-Attention hybrid architecture to obtain text with multi-dimensional feature annotations. Based on text with multi-dimensional feature annotations, it uses a bidirectional recurrent neural network to perform scene classification and dependency syntax analysis, and combines it with a dynamic knowledge graph to retrieve related entities in real time to generate structured intent data. The structured intent data is analyzed in three dimensions through the explainability engine to generate a multi-dimensional traceability report. Based on multi-dimensional traceability reports, the intelligent decision fusion engine is triggered by confidence thresholds to generate structured enhanced interaction data; Based on structured enhanced interaction data, multimodal interaction is carried out through dynamic traceability visual interface rendering and tactile coding feedback of smart wearable devices, and a complete interaction log is generated.

2. The voice interaction method based on artificial intelligence technology according to claim 1, characterized in that: The deep neural network noise reduction model is used to eliminate environmental noise. The specific steps are: A deep neural network denoising model is built based on a hybrid architecture of convolutional neural networks and recurrent neural networks. The convolutional neural network layer extracts the spectral features of the speech signal after beamforming, while the recurrent neural network layer captures the temporal features of the speech signal to distinguish noise. A deep neural network denoising model is trained using paired data of noisy and clean speech, and the difference between the denoised features and the clean speech is calculated using the mean square error loss function. The Adam optimizer is used to back-propagate and update parameters. After optimization on the validation set, the trained deep neural network denoising model is deployed to estimate and suppress the noise components. The output layer outputs the denoised speech features, which are then encoded to form a voice data packet.

3. The voice interaction method based on artificial intelligence technology according to claim 1, characterized in that: The dual-layer collaborative recognition using the main channel Transformer model and the auxiliary channel CTC-Attention hybrid architecture has the following specific steps: After preprocessing, the speech processing packet is input into the Transformer model of the main channel and the CTC-Attention model of the auxiliary channel for feature encoding; The Transformer encoding process generates a context representation through multi-head self-attention, then performs residual connection and layer normalization, and then performs residual connection and layer normalization twice after nonlinear transformation through a feedforward neural network, and finally outputs a coded feature sequence; The attention mechanism dynamically calculates the interaction weight between each character in the target text and the encoded feature sequence and generates a context vector. The time step features of the CTC-Attention model are concatenated with the attention context vector to generate an optimized feature sequence. The main channel and auxiliary channel are reconstructed into a global semantic strategy layer and a local alignment execution layer through functional decoupling. The global semantic strategy layer focuses on cross-frame semantic associations, while the local alignment execution layer strengthens the alignment confidence at the time step granularity. A gated residual fusion strategy is adopted to dynamically fuse global semantic features with local alignment features to generate multi-dimensional feature-annotated text.

4. The voice interaction method based on artificial intelligence technology according to claim 1, characterized in that: The method combines the dynamic knowledge graph to retrieve related entities in real time and generate structured intent data. The specific steps are: Extract scene classification results and dependency information from text with multi-dimensional feature annotations, dynamically combine them into query conditions, and retrieve related entities in the knowledge graph in real time; Relevant entities are filtered through entity matching and similarity calculation, and integrated into structured intent data in JSON format.

5. The voice interaction method based on artificial intelligence technology according to claim 1, characterized in that: The three-dimensional analysis is performed through the explainability engine to generate a multi-dimensional traceability report. The specific steps are: The structured intent data is input into the explainability engine. Through hierarchical feature decoupling, the structured intent data is split into three independent encoder channels: intent dimension encoder channel, entity association dimension encoder channel, and time series trajectory dimension encoder channel. The split data are processed separately. The semantic confidence vector, entity association strength vector and temporal feature vector output by the three-channel independent encoder are enhanced by residual connection and input into the Gram-Schmidt orthogonalization process to eliminate the linear correlation between dimensions and generate a mutually independent orthogonal basis vector group. A multi-dimensional traceability report is generated based on the orthogonal basis vector group.

6. The voice interaction method based on artificial intelligence technology according to claim 1, characterized in that: The intelligent decision fusion engine is triggered by the confidence threshold to generate structured enhanced interaction data. The specific steps are: Set semantic dimension thresholds, context dimension thresholds, and intent dimension thresholds, compare the confidence levels of each dimension, and dynamically trigger the intelligent decision fusion engine. The dependency annotations and associated entity annotations in the semantic dimension are fused to enhance the semantic representation vector. The scene classification results and context labels in the context dimension are context-integrated to enhance the context representation vector. The intent labels and intent verification results in the intent dimension are combined through a weighted fusion operation to optimize the intent representation vector. The fused semantic dimension, context dimension and intention dimension are integrated through data aggregation, and JSON standard encoding is used to generate structured enhanced interaction data.

7. The voice interaction method based on artificial intelligence technology according to claim 1, characterized in that: The multimodal interaction between dynamic traceability visual interface rendering and tactile coding feedback of smart wearable devices and the generation of a complete interaction log are carried out in the following specific steps: The dynamic traceability visualization interface renders a three-layer interactive view based on enhanced interaction data, and the smart wearable device tactile encoding engine converts entity association strength and intent confidence into dual-channel tactile signals; Through timestamp synchronization, tactile pulses are linked to visual animation keyframes, and pressure sensor data reversely regulates the interface mechanical parameters in real time to form a closed-loop interaction. The timeline-aligned voice data stream, visual operation path, and tactile parameter triples are encoded according to the W3CActivityStreams 2.0 specification and verified with SHA-256 to generate a complete interaction log that can be restored across devices.

8. A voice interaction system based on artificial intelligence technology, based on the voice interaction method based on artificial intelligence technology according to any one of claims 1 to 7, characterized in that: include: Acquisition module, annotation module, parsing module, analysis module, fusion module and interaction module, The acquisition module captures user voice data, synchronizes the sound source focusing algorithm to lock the target direction, and uses a deep neural network noise reduction model to eliminate environmental noise and output voice data packets; The annotation module uses a dual-layer collaborative recognition architecture consisting of a main channel Transformer model and an auxiliary channel CTC-Attention hybrid architecture to obtain text with multi-dimensional feature annotations. The parsing module, based on text with multi-dimensional feature annotations, performs scene classification and dependency syntax analysis through a bidirectional recurrent neural network, and combines it with a dynamic knowledge graph to retrieve related entities in real time to generate structured intent data; The analysis module uses the explainability engine to perform three-dimensional analysis on structured intent data and generate a multi-dimensional traceability report. The fusion module, based on multi-dimensional traceability reports, triggers the intelligent decision fusion engine through confidence thresholds to generate structured enhanced interaction data; The interaction module, based on structured enhanced interaction data, conducts multimodal interaction through dynamic traceability visual interface rendering and tactile coding feedback of smart wearable devices, and generates a complete interaction log.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the voice interaction method based on artificial intelligence technology described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the voice interaction method based on artificial intelligence technology described in any one of claims 1 to 7 are implemented.

Images