An AI Digital Human Interactive Response Method Based on a Large Language Model

Abstract

This invention discloses an AI digital human interactive response method based on a large language model, belonging to the field of digital human interaction technology. The method includes: parsing collected user voice and visual data using natural language processing methods to generate cross-modal feature vectors; performing cross-modal association parsing on the cross-modal feature vectors to generate a semantic association topology graph; calculating vertex coordinates and joint activity thresholds of the semantic association topology graph using high digital human relevance; inputting the vertex coordinates and joint activity thresholds into a constructed coordinate index library to perform attention weight calibration, outputting a multi-dimensional association graph; performing information density analysis based on the multi-dimensional association graph to generate an information density gradient vector field, and dividing it into high-density core regions and low-density edge regions. The high-density core region generates a semantic core encoding tensor, and the low-density edge region generates an edge feature encapsulation packet. This invention achieves cross-modal association modeling by converting cross-modal fusion vectors into cross-modal feature vectors.

Description

Technical Field

[0001] This invention relates to the field of digital human interaction technology, and in particular to an AI digital human interactive response method based on a large language model. Background Technology

[0002] In recent years, AI-driven digital human interaction technologies have made significant progress in the fields of multimodal perception and response generation. Deep learning-based dialogue systems achieve semantic understanding and contextual coherence through pre-trained language models, and combine computer vision technology to synchronously generate facial expressions and body movements. Multimodal fusion architectures are gradually shifting from feature-level concatenation to attention-driven cross-modal alignment, significantly improving the ability to jointly analyze user speech, facial expressions, and gestures. The industry has already achieved initial deployments of digital humans in customer service, equipment monitoring, and other scenarios. Their core relies on end-to-end neural network architectures, achieving collaborative optimization of speech, vision, and control signals through joint training.

[0003] Current AI-powered digital human interactive response methods have several shortcomings. First, they rely on manual feature engineering or static attention weight allocation, making it difficult to dynamically capture the nonlinear entanglement relationships between multi-source data. Implicit cross-modal correlations in complex scenes are simplified to linear correlation analysis, resulting in a significant increase in semantic parsing errors. Second, response generation lacks embedded constraints based on physical laws. The conversion from natural language responses to control commands uses independent training modules, and no spatiotemporal consistency guarantee mechanism based on differential geometry has been established, leading to millisecond-level coordination deviations between motion trajectories and acoustic-optical responses. Summary of the Invention

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

[0005] Therefore, this invention provides an AI digital human interactive response method based on a large language model to solve the problems of insufficient cross-modal association and lack of physical constraints.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides an AI digital human interactive response method based on a large language model. The method includes: parsing collected user voice and visual data using natural language processing methods to generate cross-modal feature vectors; performing cross-modal association parsing on the cross-modal feature vectors to generate a semantic association topology graph; calculating vertex coordinates and joint activity thresholds of the semantic association topology graph using high digital human relevance; inputting the vertex coordinates and joint activity thresholds into a constructed coordinate index library to perform attention weight calibration and output a multidimensional association graph; performing information density analysis based on the multidimensional association graph to generate an information density gradient vector field, and dividing it into a high-density core region and a low-density edge region; generating a semantic core encoding tensor in the high-density core region and an edge feature encapsulation package in the low-density edge region; constructing a dynamically optimized language model, inputting the semantic core encoding tensor and the edge feature encapsulation package to generate a natural language text response, and converting it into a multimodal response instruction set; extracting control parameters from the multimodal response instruction set, generating corresponding synchronized audio streams, facial expressions, and limb movement trajectories using the control parameters, and integrating the synchronized audio streams, facial expressions, and limb movement trajectories to generate interactive response behavior.

[0008] As a preferred embodiment of the AI ​​digital human interactive response method based on a large language model described in this invention, the specific steps for generating the semantic association topology graph are as follows:

[0009] User voice data is used to generate word-segmented text sequences through speech recognition, and visual data is used to extract visual feature vectors through convolution calculation.

[0010] Word embedding is performed on the segmented text sequence to generate text vectors, which are then input together with the visual feature vectors into a sparse attention mechanism to generate cross-modal fusion vectors.

[0011] Cross-modal fusion vectors generate cross-modal feature vectors through feature space projection; cross-modal feature vectors construct semantic association topology graphs through cross-modal association parsing.

[0012] As a preferred embodiment of the AI ​​digital human interactive response method based on a large language model described in this invention, the specific steps for calculating the vertex coordinates and joint activity thresholds of the semantic association topology graph through high digital human relevance are as follows:

[0013] Vertex and edge weight data of the semantic association topology graph are used to generate joint motion constraint tensors through high digital human relevance;

[0014] The vertex positions are dynamically adjusted based on the joint motion constraint tensor, and the joint activity thresholds corresponding to the edge weights are mapped to output the vertex coordinates and joint activity thresholds.

[0015] As a preferred embodiment of the AI ​​digital human interactive response method based on a large language model described in this invention, the specific steps for outputting a multidimensional association graph are as follows:

[0016] Spatial encoding of vertex coordinates, association of joint activity thresholds with corresponding vertex pairs to form a coordinate-joint mapping table, and injection of timestamps and information entropy labels to build a coordinate index library;

[0017] The coordinates of neighboring vertices are retrieved by spatial indexing, the joint activity threshold is obtained by calling the coordinate-joint mapping table, and the association strength threshold function is constructed. The association transition probability is calculated. If the association transition probability is greater than the preset probability threshold, a multidimensional association map is generated.

[0018] As a preferred embodiment of the AI ​​digital human interactive response method based on a large language model described in this invention, the specific steps for dividing the high-density core region and the low-density edge region are as follows:

[0019] Input the multidimensional correlation map into the information density analysis formula to generate an information density distribution map. Perform gradient operator operations on the information density distribution map to generate an information density gradient vector field.

[0020] The information density gradient magnitude is calculated based on the information density gradient vector field and compared with a preset entropy threshold. If the information density gradient magnitude exceeds the entropy threshold, it is classified as a high-density core region; otherwise, it is classified as a low-density edge region.

[0021] As a preferred embodiment of the AI ​​digital human interactive response method based on a large language model described in this invention, the high-density core region generates a semantic core encoding tensor, and the low-density edge region generates an edge feature encapsulation packet. The specific steps are as follows:

[0022] The high-density core region is simplified through graph structure and compressed by tensor decomposition to generate a semantic core encoding tensor;

[0023] Low-density edge regions are processed by edge semantic entropy compression calculation and edge feature encoding to generate edge feature encapsulation packages.

[0024] As a preferred embodiment of the AI ​​digital human interactive response method based on a large language model described in this invention, the specific steps for constructing the dynamically optimized language model are as follows:

[0025] A decoding and fusion layer is constructed based on the principle of topology reconstruction; a response optimization layer is constructed based on the principle of parameter mapping; and a language generation layer is constructed based on the principle of template filling.

[0026] A dynamically optimized language model is constructed through a decoding fusion layer, a response optimization layer, and a language generation layer.

[0027] As a preferred embodiment of the AI ​​digital human interactive response method based on a large language model described in this invention, the specific steps of converting to a multimodal response instruction set are as follows:

[0028] The semantic core encoding tensor and edge feature encapsulation package are input into the dynamically optimized language model, and natural language text response is generated through topological reconstruction, parameter mapping and template filling.

[0029] Transform natural language text responses into multimodal response instruction sets through cross-modal instruction conversion.

[0030] As a preferred embodiment of the AI ​​digital human interactive response method based on a large language model described in this invention, the control parameters include voice parameters, facial expression parameters, and action parameters.

[0031] As a preferred embodiment of the AI ​​digital human interactive response method based on a large language model described in this invention, the specific steps for generating interactive response behavior are as follows:

[0032] Speech parameters are used to generate a synchronized audio stream through Mel spectrum generation and vocoder conversion;

[0033] Facial expressions are generated by simulating facial muscles using facial expression parameters.

[0034] Motion parameters are solved using inverse kinematics to generate limb motion trajectories;

[0035] By aligning synchronized audio streams, facial expressions, and body movement trajectories with timestamps and fusing them with behaviors, interactive response behaviors are generated.

[0036] The beneficial effects of this invention are as follows: by projecting the cross-modal fusion vector into a cross-modal feature vector through feature space, a semantic association topology graph is constructed, which realizes accurate modeling of cross-modal association relationships and significantly reduces semantic parsing errors; by dynamically optimizing the vertex distribution of the semantic association topology graph and combining the joint motion constraint tensor to calculate the joint activity threshold, physical constraints are embedded into the response generation link, realizing differential geometrical coordination between motion trajectory and acoustic-optical response, and completely eliminating the millisecond-level delay generated by independent modules. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a flowchart of an AI digital human interactive response method based on a large language model.

[0039] Figure 2 A flowchart for generating a multidimensional correlation graph.

[0040] Figure 3 A flowchart for constructing a dynamically optimized language model.

[0041] Figure 4 A flowchart for generating interactive response behavior. Detailed Implementation

[0042] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0043] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0044] Secondly, the term "one embodiment" or "embodiment" as used 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 different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0045] Reference Figures 1-4 This is one embodiment of the present invention, which provides an AI digital human interactive response method based on a large language model, including the following steps:

[0046] S1. Parse the collected user voice and visual data using natural language processing methods to generate cross-modal feature vectors. Then, perform cross-modal association parsing on the cross-modal feature vectors to generate a semantic association topology graph.

[0047] S1.1 User voice data is used to generate word-segmented text sequences through speech recognition, and visual data is used to extract visual feature vectors through convolution calculation;

[0048] It should be noted that the user speech data undergoes frame segmentation, with each frame being 25 milliseconds long. A Hamming window function is applied to each frame of user speech data to reduce spectral leakage. The weighted user speech data is then subjected to a Fast Fourier Transform (FFT) to obtain a complex spectrum. This complex spectrum is input into a Mel-scale triangular filter bank to obtain the energy values ​​for each frequency band. After logarithmic compression, the frequency band energy values ​​are subjected to a Discrete Cosine Transform (DCT) to obtain Mel-frequency cepstral coefficients, which serve as acoustic features. These acoustic features are then input into a Hidden Markov State Transition Probability Matrix (HMM), and the probability distribution of each acoustic feature is calculated using the Viterbi algorithm to obtain the phoneme sequence of the most probable path. The phoneme sequence is then matched against a predefined dictionary to match Chinese syllable combinations. Combined with binary grammar probability, the word segmentation boundaries are adjusted to generate a segmented text sequence. The expression for the Viterbi algorithm is...

[0049] lnδ t (j)=max i [lnδ t-1 (i)+lna ij ]+ln b j (o t );

[0050] Where, δ t (j) represents the factor sequence with the maximum cumulative probability of reaching state j at frame t; i represents the phoneme state index of the previous frame, for example, i = 2 = / i / sound; j represents the phoneme state index of the current frame, for example, j = 3 = / sh / sound; t represents the frame number; a ij This represents the transition probability of a phoneme state in the Hidden Markov state transition probability matrix, such as the probability of transitioning from the / sh / sound to the / i / sound; b j (·) represents the probability density function for generating observations from state j; t Represents the observation value at frame t; max i This means iterating through all phoneme states to find the largest value of i;

[0051] Visual data is processed by sliding multi-layer convolutional kernels to extract spatial features. Each convolutional kernel performs ReLU activation to enhance nonlinear representation. The numerical distribution is stabilized by normalization, and global average pooling is performed to compress it into a visual feature vector.

[0052] It should also be noted that the predefined dictionary is formed by collecting Chinese text corpora in the target domain, marking Chinese vocabulary boundaries and extracting high-frequency terms, merging synonyms and abbreviations to form a basic vocabulary list; adding words from standard Chinese dictionaries to the basic vocabulary list, covering commonly used verbs, nouns and professional terms; performing binary grammatical relation statistics on the added basic vocabulary list to calculate word frequency co-occurrence probability; reviewing and correcting ambiguous word boundaries, constructing a syllable-to-Chinese character mapping table and a vocabulary probability table, and storing them as a predefined dictionary.

[0053] S1.2 Perform word embedding operation on the segmented text sequence to generate text vectors, and input them together with the visual feature vectors into the sparse attention mechanism to generate cross-modal fusion vectors;

[0054] It should be noted that each word in the segmented text sequence is converted into a 300-dimensional list of values ​​as a text vector through pre-trained word vector mapping; the visual feature vector is compressed to 256 dimensions through global average pooling, and then mapped to 300 dimensions through a fully connected layer, aligned with the text vector; the text vector and visual feature vector are concatenated into a joint matrix and evenly divided into eight groups of sub-vectors, each group of sub-vectors undergoing independent self-attention calculation: the query vector and key vector of the current group are extracted, a dot product is performed, and the ratio of the dot product result to the square root of the feature dimension is calculated, and then an exponential function is applied. The mathematical processing of functions and summation normalization is transformed into a probability weight distribution. The probability weight distribution is then element-wise weighted and summed with the value vectors of the same group to obtain a single attention output vector. Eight single attention output vectors are concatenated along the feature dimensions to form a 512-dimensional high-dimensional composite vector. Feature filtering is performed on this 512-dimensional composite vector, calculating the association strength between each feature dimension and the target task. Only high-value feature dimensions with an association strength exceeding 0.7 (based on the statistical definition of cross-modal feature association strength distribution in the target domain) are retained, while the remaining dimensions are set to zero, generating a cross-modal fusion vector. For example, when processing the voice command "abnormal device vibration," the text vector of the word "vibration" will form a strong association with the feature vector of the high-frequency flickering area in the visual features during attention calculation, while irrelevant background features are sparsified and filtered.

[0055] It should also be noted that the pre-trained word vectors are generated by collecting large-scale general text corpora (such as Wikipedia, news web pages, and books and documents), statistically analyzing the frequency of word occurrence and the distribution of co-occurring words in context, and constructing a word-context association matrix. The word-context association matrix is ​​then subjected to dimensionality reduction decomposition to extract semantic features, mapping each word to a 300-dimensional real vector space. Words with similar semantics are located close to each other in the vector space, forming a fixed mapping table between words and 300-dimensional vectors, which is the word-vector mapping table. The word-vector mapping table is further injected with vectors of target domain professional terms to form a fused word vector set containing general vocabulary and target domain professional terms, which is stored as pre-trained word vectors.

[0056] S1.3 Cross-modal fusion vectors generate cross-modal feature vectors through feature space projection; cross-modal feature vectors construct semantic association topology graphs through cross-modal association parsing.

[0057] It should be noted that the cross-modal fusion vector is expanded using Fourier orthogonal basis functions. Each component of the 512-dimensional cross-modal fusion vector is multiplied by 1024 orthogonal basis functions. The orthogonal basis functions cover the phase of the entire frequency domain in complex exponential form. Each inner product result generates a complex coordinate value. The expansion results of all components are combined into a 1024-dimensional complex vector, which is the cross-modal feature vector.

[0058] The cross-modal feature vector contains 1024 complex elements (each element is a+b). i The algorithm iterates through all element pairs (e.g., elements A and B, A and C, etc.), performs a complex inner product operation on each element pair, and takes the square of the modulus value, which is the association strength value between the two elements. If the association strength value is greater than the preset association strength threshold (based on the definition of the critical value of cross-modal feature vector association strength in the target domain, the 85th percentile value of the cumulative probability distribution of cross-modal feature vector association strength in historical data of the target domain, such as 0.65), then a vertex connection edge is established, where the vertex is a cross-modal feature vector element, the edge weight is the association strength value, and the output is a semantic association topology graph with weights and a directed structure.

[0059] S2. Calculate the vertex coordinates and joint activity thresholds of the semantic association topology graph through high digital human relevance, input the vertex coordinates and joint activity thresholds into the constructed coordinate index library to perform attention weight calibration, and output a multidimensional association graph.

[0060] S2.1 The vertex data and edge weight data of the semantic association topology graph are used to generate a joint motion constraint tensor through high digital human relevance;

[0061] It should be noted that converting the data of each vertex in the semantic association topology graph into three-dimensional spatial coordinates of the digital human skeleton joints involves the following steps: The original two-dimensional coordinate data of each vertex (including the values ​​of the x-coordinate and y-coordinate) is normalized to the 0-1 range; a three-dimensional mapping is performed based on the spatial distribution of the digital human skeleton joints: the product of the normalized x-coordinate and the digital human height ratio coefficient (based on the statistical definition of the ratio of shoulder width to height in adult males in anthropometry database, such as 0.8) is used as the horizontal displacement of the joint; the product of the normalized y-coordinate and the human body thickness parameter (based on the maximum definition of the thickness of the human coronal trunk, such as 25cm) is used as the vertical displacement of the joint; and coordinate superposition is performed using the skeletal reference point (based on biomechanical characteristics): the horizontal displacement is superimposed on the x-axis coordinate, the vertical displacement is superimposed on the z-axis coordinate, and the y-axis coordinate of the reference point remains unchanged, generating complete three-dimensional joint coordinates.

[0062] The edge weights are mapped to the range of joint movement angles. Specifically, the product of the edge weights and the maximum physiological movement angle of the corresponding joints in the human body is used as the basic angle range. Then, joint type compensation coefficients (based on the virtual skeleton collision detection experiment definition in the target domain, such as shoulder joint compensation coefficient 0.9 and wrist joint compensation coefficient 1.1) are superimposed for biomechanical adaptation.

[0063] The joint motion constraint tensor is calculated using the ergonomic constraint formula, and its expression is:

[0064] T = u·C·v;

[0065] Where T represents the joint motion constraint tensor; u represents the basic angle range of the joint; C represents the joint type compensation coefficient; and v represents the safety margin correction term, based on the definition of industrial robot safety specifications.

[0066] S2.2. Dynamically adjust vertex positions based on joint motion constraint tensors, map joint activity thresholds corresponding to edge weights, and output vertex coordinates and joint activity thresholds;

[0067] It should be noted that the three-dimensional joint coordinates in the joint motion constraint tensor are used as the initial position input, and the vertex position is dynamically adjusted based on ergonomic constraints: the basic joint angle range stored in the joint motion constraint tensor is extracted, and the angular deviation value between the current joint vector and the target direction is calculated; if the angular deviation value exceeds the deviation threshold (based on the definition of joint motion deviation tolerance, such as 0.5°), the joint coordinates are moved along the normal vector direction, and the movement step size is determined by the product of the angular deviation value and the biomechanical elastic coefficient; the joint position is iteratively adjusted until all angular deviations are below the 2° tolerance (based on the definition of human joint activity measurement method), and the optimized three-dimensional joint coordinates are output as vertex position data; at the same time, the basic joint angle range in the joint motion constraint tensor is directly mapped to the joint activity threshold corresponding to the edge weight, generating vertex coordinates and joint activity thresholds.

[0068] S2.3. Spatial encoding of vertex coordinates, association of joint activity thresholds with corresponding vertex pairs to form a coordinate-joint mapping table, and injection of timestamps and information entropy labels to build a coordinate index library;

[0069] It should be noted that the x-coordinate and y-coordinate values ​​of the vertex coordinates are converted into binary digit sequences respectively. The binary digit sequences of the x-coordinate and y-coordinate are then rearranged and combined according to an interleaving rule. The least significant bit of the x-coordinate is inserted alternately with the least significant bit of the y-coordinate, and the second least significant bit of the x-coordinate is inserted alternately with the second least significant bit of the y-coordinate, until all digits are processed, forming a mixed binary digit sequence. The mixed binary digit sequence is then converted into a decimal integer value, which is the spatial index value.

[0070] The starting point spatial index value and the ending point spatial index value are combined to form a vertex pair key, and the joint activity threshold is used as the associated value. The vertex pair key and the joint activity threshold form a one-to-one key-value relationship. The set of vertex pair keys and joint activity thresholds of all edges constitutes a complete mapping table, forming a coordinate-joint mapping table.

[0071] The coordinate-joint mapping table adds the current time value as a timestamp, calculates the information entropy value of the vertex coordinates as a stability label, and integrates the spatial index value, joint activity threshold, timestamp value and stability label into a structured record to build a coordinate index library.

[0072] The expression for calculating the information entropy value is:

[0073]

[0074] Where H represents the information entropy value; z represents the spatial region index, with a value of 1-n; n represents the total number of spatial regions; p z This represents the probability that the vertex coordinates fall within the z-th spatial region.

[0075] S2.4. Retrieve the coordinates of neighboring vertices through spatial index, call the coordinate-joint mapping table to obtain the joint activity threshold and construct the association strength threshold function, calculate the association transition probability, and if the association transition probability is greater than the preset probability threshold, generate a multidimensional association map.

[0076] It should be noted that the probability threshold is based on the determination of the background noise level of the computing hardware. The correlation transition probability distribution is measured by calibration experiment when there is no signal input, and the 99th percentile value of the correlation transition probability distribution is taken as the preset probability threshold.

[0077] Continuous proximity retrieval is triggered by the spatial index value of the target vertex. The indexes of neighboring vertices that are continuously adjacent to the spatial index value of the target vertex are extracted from the coordinate index library. The target vertex pair key is formed based on the target index and the neighboring vertex index, and the joint activity threshold is obtained by querying the coordinate-joint mapping table through the target vertex pair key. The association strength threshold function is constructed by the reciprocal of the joint activity threshold, and the association transition probability is calculated by inputting the association transition probability formula. If the association transition probability is greater than the probability threshold, the coordinate values ​​of neighboring vertices, the joint activity threshold, and the stability label are extracted from the coordinate index library to form a multidimensional association map.

[0078] The expression for the association strength threshold function is:

[0079]

[0080] Where V(·) represents the correlation strength threshold function; θ represents the distance from the current position to the starting point, which is a mathematical variable of spatial position used to quantify the relative position of the target object in the state change path, for example: imagine a straight path from the starting point to the ending point (such as the gradual process of equipment changing from normal to fault); v represents the joint activity threshold; A represents the barrier strength constant, which is defined based on the target domain state transition resistance benchmark.

[0081] The formula for calculating the probability of a correlated transition is as follows:

[0082] P = e -E·θ ;

[0083] Where P represents the probability of association transition; E represents the scaling factor, defined based on prior knowledge of the semantic association strength in the target domain.

[0084] It should also be noted that existing technologies achieve basic association modeling through artificial feature engineering combined with static manifold learning, thus achieving data dimensionality reduction. However, they still have fundamental limitations: they cannot resolve feature phase consistency, curvature calculation is separated from association strength, and fault identification relies on human experience. This solution achieves deep fusion of geometry and association: the separation of real and imaginary parts of features preserves phase consistency; the inverse proportional conversion of association strength to geodesic distance breaks through the limitations of Euclidean space; and the reciprocal of joint activity threshold dynamically constructs an information density field, improving recognition speed.

[0085] S3. Perform information density analysis based on the multidimensional correlation graph, generate information density gradient vector field, and divide the high-density core area and low-density edge area. Generate semantic core encoding tensor in the high-density core area and generate edge feature encapsulation package in the low-density edge area.

[0086] S3.1 Input the multidimensional correlation map into the information density analysis formula to generate an information density distribution map. Perform gradient operator operations on the information density distribution map to generate an information density gradient vector field.

[0087] It should be noted that the coordinate values ​​of neighboring vertices and the joint activity threshold are extracted from the multidimensional correlation graph. Based on the coordinate values ​​of neighboring vertices, a triangular meshing operation is performed: neighboring vertices are connected to form non-intersecting triangular structures, ensuring that there are no other vertices inside the circumcircle of all triangles, forming a mesh network of seamlessly spliced ​​triangular faces. Each triangle has three edges connecting to neighboring vertices, and the triangular faces share edges and do not overlap.

[0088] The arithmetic mean of the reciprocals of the joint activity thresholds corresponding to each vertex is taken as the information density value; the information density values ​​of all vertices are combined to generate a two-dimensional information density distribution map; the Sobel gradient operator convolution operation is performed on the two-dimensional information density distribution map; the partial derivative components of the information density values ​​in the horizontal and vertical coordinate directions are calculated respectively, and the expression is:

[0089]

[0090] Among them, G d (x, y) represents the partial derivative components; d represents the direction of the partial derivative components. When d = x, G d (x, y) represents the partial derivative components in the horizontal direction. When d = y, G d (x,y) represents the partial derivative components along the vertical axis; D represents the two-dimensional information density distribution map; M d Represents the horizontal and vertical convolution kernels; ω represents the vertical convolution kernel offset, with a value range of {-1, 0, 1}; ω represents the horizontal convolution kernel offset, with a value range of {-1, 0, 1}.

[0091] It should also be noted that the Sobel gradient operator is a 3×3 convolution kernel, which requires traversing the 8 neighboring pixels (top / bottom / left / right / top-left / top-right / bottom-left / bottom-right) around the target position (x,y). The nine combinations of ω cover the entire neighborhood, ensuring that gradient calculation fully captures the local variations of information density values ​​in both the vertical and horizontal directions. The nine combinations are:

[0092]

[0093] The horizontal and vertical partial derivative components of each vertex are combined to form the information density gradient vector, and the magnitude of the information density gradient is calculated, expressed as follows:

[0094]

[0095] in, Represents the information density gradient vector; G represents the magnitude of the information density gradient. x G represents the partial derivative components along the horizontal axis; y The partial derivative components in the ordinate direction;

[0096] The information density gradient vectors of all vertices constitute the information density gradient vector field. The vector direction represents the direction of maximum increase in information density, and the magnitude of the information density gradient represents the intensity of density change.

[0097] S3.2 Compare the information density gradient magnitude with the preset entropy threshold; if the information density gradient magnitude exceeds the entropy threshold, it is classified as a high-density core region, otherwise it is classified as a low-density edge region.

[0098] It should be noted that the entropy threshold is dynamically calculated using an exponential decay function, expressed as follows:

[0099] τ=α·e -βH ;

[0100] Where τ represents the entropy threshold; α represents the domain density benchmark value, defined based on the physical or application characteristics of the target domain; β represents the information attenuation constant, defined based on the propagation law of electromagnetic waves in a medium, with a fixed value of 0.5; and H represents the information entropy value.

[0101] If the magnitude of the information density gradient is greater than the entropy threshold, the information density gradient vector field region where the corresponding vertex is located is marked as a high-density core region; otherwise, it is marked as a low-density edge region. The high-density core region represents a region with drastic changes in information intensity, while the low-density edge region represents a region with gentle information flow.

[0102] S3.3. The high-density core region generates a semantic core encoding tensor through graph structure simplification and tensor decomposition compression; the low-density edge region generates an edge feature encapsulation package through edge semantic entropy compression calculation and edge feature encoding.

[0103] It should be noted that the high-density core region uses the reciprocal of the joint activity threshold as the association strength index. Edges with an association strength index greater than the association threshold (defined based on the conductivity characteristics of the target domain material) are defined as strong core connections, and edges with an association strength index less than the association threshold are defined as weak secondary connections. All weak secondary connections are removed, and only strong core connections are retained to form a simplified topology network.

[0104] In the simplified topology network, each vertex contains two feature components: vertex coordinates and correlation strength index. These two feature components are projected onto mutually orthogonal coordinate axes to form independent feature dimensions. The vertex coordinates are mapped to horizontal axis scale intervals, and each scale interval is assigned a unique binary address code. The address code of the interval containing the vertex coordinates serves as the position encoding sequence. The correlation strength index is projected onto the vertical axis to generate an intensity encoding sequence. The correlation strength index is segmented into levels based on the intensity range boundary values ​​(defined by the conductivity properties of the target domain material). After the correlation strength index falls into the corresponding level, the corresponding level number is converted into a binary intensity level code, which serves as the intensity encoding sequence. The position encoding sequence and the intensity encoding sequence are combined according to spatial orthogonality to generate a semantic core encoding tensor.

[0105] The information density value is used as the diagonal element of the density matrix, and the off-diagonal elements of the density matrix are set to zero to generate the density matrix. The edge semantic entropy compression value is obtained by the logarithmic operation of the diagonal elements of the density matrix. The product of the edge semantic entropy compression value and the constant of pi is used as the phase angle. The sine function value of the phase angle is used as the imaginary amplitude, and the cosine function value is used as the real amplitude. The real and imaginary amplitudes are combined to form a complex amplitude value. The complex amplitude value is converted into a text string format, and the real and imaginary values ​​are retained to three decimal places. The phase angle is retained in radians to generate an edge feature encapsulation package.

[0106] It should also be noted that existing technologies achieve data compression through uniform sampling or fixed threshold segmentation, simplifying basic information, but still have fundamental limitations: they cannot distinguish information value density, and high-value features and noise are compressed equally, leading to the loss of key information; static compression rules are difficult to adapt to dynamic data fluctuations, and edge information is directly discarded, causing contextual breaks. This solution achieves value-aware dynamic compression: the information density gradient field accurately divides the core area and the edge area; the core area retains strongly correlated topological features to generate a semantic core encoding tensor; the edge area retains phase coherence through dynamic correlation entropy encapsulation; the dynamic entropy threshold adapts to the data distribution, improving the integrity of fault features.

[0107] S4. Construct a dynamically optimized language model, input the semantic core encoding tensor and edge feature encapsulation package, generate a natural language text response, and convert it into a multimodal response instruction set;

[0108] S4.1 Construct a decoding and fusion layer based on the principle of topology reconstruction, a response optimization layer based on the principle of parameter mapping, and a language generation layer based on the principle of template filling;

[0109] It should be noted that the system includes a coordinate analysis component, an intensity analysis component, and an amplitude and phase analysis component. The coordinate analysis component groups the position encoding sequence into binary bits, maps each group of bits to a scale interval, and takes the median of the scale interval to restore the geometric coordinate value. The intensity analysis component converts the intensity encoding sequence into a decimal level number, with the level number corresponding to the boundary value of the intensity range. The median of the boundary value of the intensity range is taken to restore the associated intensity value. The amplitude and phase analysis component analyzes the edge feature encapsulation text string, extracts the complex amplitude value string and converts it into a floating-point amplitude value, and extracts the phase angle value string and converts it into a floating-point phase angle value, thus restoring the complex amplitude value and phase angle value. The geometric coordinate value is projected onto the spatial semantic axis, the associated intensity value is projected onto the associated semantic axis, the complex amplitude value is projected onto the wave semantic axis, and the phase angle value is projected onto the stable semantic axis. The four semantic axes are orthogonally combined to form a fused semantic space, thus completing the construction of the decoding fusion layer.

[0110] The system is configured with a spatial coordinate projection component, a correlation intensity projection component, a fluctuation amplitude projection component, and a stable phase projection component. The spatial coordinate projection component outputs geometric coordinate values ​​directly as spatial coordinate parameters. The correlation intensity projection component uses the product of the correlation intensity value and the conversion coefficient (defined based on the physical conduction efficiency and risk buffer ratio of the target domain) as the coupling intensity parameter. The fluctuation amplitude projection component uses the magnitude of the complex amplitude value as the magnetic field strength parameter. The stable phase projection component uses the ratio of the phase angle value to the constant pi as the time parameter. The four types of parameters are combined to form the response path parameter set, thus completing the construction of the response optimization layer.

[0111] The system sets up a path parameter receiving interface, a semantic extraction component, and a template filling component. The path parameter receiving interface acquires the spatial coordinate parameters, coupling strength parameters, magnetic field strength parameters, and time parameters output by the response optimization layer. The semantic extraction component retains two decimal places for the spatial coordinate parameters according to the precision scale (based on the target domain positioning accuracy requirements) and adds coordinate axis direction descriptors to generate a positional semantic description (e.g., at 0.70 on the X-axis). The coupling strength parameter matches the strength range boundary values ​​with level labels to generate a relational strength semantic description (e.g., a strength value of 1.18 matches "high" level → "high vibration conduction"). The magnetic field strength parameter is matched with the boundary value of the impact level (based on the standard definition of the safety threshold in the target domain) to generate an impact strength semantic description (e.g., magnetic field strength 1.0 matches "high risk" level → "high risk diffusion"); the template filling component fills the location semantic description into the location description template slot, the association strength semantic description into the relationship description template slot, and the impact strength semantic description into the impact description template slot, and combines them to generate a basic response sentence; the timeliness adjustment component is set and the time parameter is input, and the timeliness expression of the response sentence is optimized according to the time parameter value, and the natural language text response is output, completing the construction of the language generation layer.

[0112] S4.2 Construct a dynamically optimized language model through a decoding fusion layer, a response optimization layer, and a language generation layer;

[0113] It should be noted that the decoding fusion layer, response optimization layer, and language generation layer are connected in series to construct a dynamically optimized language model;

[0114] The physical parameters such as vibration, temperature, and current collected in real time from the device sensor network in the target field, as well as the operation report text recorded in the maintenance log, are combined to form a historical record database of the target field;

[0115] Input and output samples are obtained from a historical database of the target domain. The input sample is a historical semantic core encoding tensor and a historical edge feature encapsulation package, and the output sample is the corresponding historical natural language text response. During training, the input sample is processed by a decoding fusion layer to generate a fused semantic space, the response optimization layer outputs the response path parameter set, and the language generation layer synthesizes the training natural language text response. The semantic difference loss value between the training natural language text response and the historical natural language text response of the output sample is calculated. The semantic difference loss value is used to update the dynamic optimization language model parameters through gradient backpropagation. The semantic difference loss value is calculated using cosine similarity and word shift distance weighting. The gradient update adopts an adaptive momentum optimization algorithm. The initial learning rate is set to 0.001 (based on the convergence speed requirement of typical optimizers in the target domain) and decays exponentially with the training rounds. The decay coefficient is 0.95 (based on the empirical value of convergence characteristics of typical optimizers). The training termination condition is that the semantic difference loss value decreases by less than 0.0001 for ten consecutive rounds (based on the semantic error tolerance definition of the target domain) or the maximum training round of 1000 rounds is reached, thus completing the training of the dynamically optimized language model.

[0116] S4.3 Input the semantic core encoding tensor and edge feature encapsulation package into the dynamically optimized language model, and generate natural language text response through topological reconstruction, parameter mapping and template filling;

[0117] It should be noted that the position encoding sequence of the semantic core encoding tensor is restored to geometric coordinate values, and the intensity encoding sequence is restored to association intensity values; the edge feature encapsulation packet input is parsed into complex amplitude values ​​and phase angle values; the geometric coordinate values ​​are projected onto the spatial semantic axis, the association intensity values ​​are projected onto the association semantic axis, the complex amplitude values ​​are projected onto the wave semantic axis, and the phase angle values ​​are projected onto the stable semantic axis, generating a fused semantic space with four orthogonal axes; in the fused semantic space, the geometric coordinate values ​​are output as spatial coordinate parameters, the association intensity values ​​are converted into coupling intensity parameters, the complex amplitude values ​​are converted into magnetic field intensity parameters, and the phase angle values ​​are converted into time parameters; a positional semantic description is generated based on the spatial coordinate parameters, an association intensity semantic description is generated based on the coupling intensity parameters, and an influence intensity semantic description is generated based on the magnetic field intensity parameters; the positional semantic description, the association intensity semantic description, and the influence intensity semantic description are filled into the corresponding positions of the template slots to combine the basic response sentence; the time parameter optimizes the timeliness expression of the response sentence, and the natural language text response is output.

[0118] S4.4. Convert natural language text responses into multimodal response instruction sets through cross-modal instruction conversion.

[0119] It should be noted that four types of semantic elements are extracted from the natural language text response: location semantic description, association strength semantic description, influence strength semantic description, and time parameter. The location semantic description extracts coordinate axis identifiers and coordinate values ​​through character segmentation. The coordinate axis identifiers are matched with the spatial coordinate system (based on the target domain's device motion dimension definition) to generate axis type instructions, and the coordinate values ​​retain the precision scale to output coordinate value instructions, which are then combined into spatial coordinate instructions. The association strength semantic description generates intensity level instructions by matching level labels (based on the target domain's safety operation standard definition). The influence strength semantic description generates alarm instructions by matching influence labels (based on the target domain's risk threshold standard definition). The product of the time parameter value and the unit conversion factor (based on the target domain's time unit requirement definition) is converted into seconds, and the seconds are output as a countdown instruction.

[0120] Spatial coordinate commands are mapped to the device positioning interface to generate robotic arm movement commands; intensity level commands are mapped to the actuator control interface to generate vibration suppression parameters; alarm commands are mapped to the human-machine interface to generate audible and visual alarm parameters; and countdown commands are mapped to the task scheduling interface to generate maintenance queue priorities. Combining robotic arm movement commands, vibration suppression parameters, audible and visual alarm parameters, and maintenance queue priorities, a multimodal response command set is generated.

[0121] It should also be noted that existing technologies rely on manual rule templates, resulting in high costs for cross-domain adaptation and delayed response times; traditional natural language models cannot parse physical parameters such as joint activity thresholds, leading to high errors in industrial fault location; multimodal conversion systems require independent training of visual / control / speech models, resulting in high command coordination latency; this solution achieves end-to-end conversion of topology data to multimodal commands through dynamic optimization of the language model, completing deep integration of features and domain semantics; topology compression encoding preserves associated joint activity thresholds, semantic axis orthogonal projection is compatible with target domain parameters, dynamic generation of cross-modal commands improves mechanical response speed, and time parameter-driven timeliness optimization overcomes risk control latency.

[0122] S5. Extract control parameters from the multimodal response instruction set, generate corresponding synchronous audio streams, facial expressions, and limb movement trajectories through the control parameters, integrate the synchronous audio streams, facial expressions, and limb movement trajectories, and generate interactive response behavior.

[0123] S5.1 Extract speech parameters, facial expression parameters, and motion parameters from the multimodal response instruction set;

[0124] It should be noted that the target axis identifier and displacement coordinate value are extracted from the robotic arm movement commands contained in the multimodal response instruction set, and the vibration suppression intensity percentage value is extracted from the vibration suppression parameters. The target axis identifier, displacement coordinate value, and vibration suppression intensity percentage value are combined to construct the motion parameters. The audio file identifier and optical attribute label are separated from the audible and visual alarm parameters. The audio file identifier generates a voice index, and the optical attribute label is parsed into RGB color values ​​and flashing frequency to form expression parameters. The voice index is directly output as the voice parameters.

[0125] S5.2. Speech parameters are generated into a synchronous audio stream through Mel spectrum generation and vocoder conversion; facial expression parameters are generated into facial expressions through facial muscle simulation; and motion parameters are generated into limb motion trajectories through inverse kinematics solution.

[0126] It should be noted that the audio resources are a collection of audio files constructed by collecting typical acoustic samples from the target domain. The audio file collection is stored as a waveform data file library in pulse code modulation format. When the speech parameter is an audio file identifier, the audio resources are called to directly decode into a pulse code modulation digital waveform.

[0127] When the speech parameters are text content, they are converted into a short-time spectrum through time-domain waveform framing and windowing. A Mel-scale triangular filter bank is applied to the short-time spectrum to extract the energy values ​​of each frequency band, generating a Mel-frequency cepstral coefficient feature matrix. This matrix is ​​then input into a vocoder to establish vocal tract formants and superimposed with a fundamental frequency excitation source to synthesize a pulse code modulation digital waveform. The pulse code modulation digital waveform is input into a digital-to-analog converter (DAC). The DAC reads the voltage amplitude values ​​at each discrete time point in sequence according to the sampling rate. The voltage amplitude values ​​are converted into a continuous stepped voltage signal through a resistor network stepped circuit. The continuous stepped voltage signal is smoothed by a low-pass filter to remove high-frequency quantization noise, and the continuously changing analog voltage signal output is the synchronous audio stream.

[0128] The fundamental frequency excitation source is generated by the vocoder extracting the fundamental frequency parameters from the periodic components of the Mel frequency cepstral coefficient feature matrix. The fundamental frequency parameters control the generation of the fundamental frequency excitation source by the periodic pulse sequence.

[0129] The facial expression parameters are input into the facial driving operation. If the expression parameters include RGB color values, the color mapping relationship of the facial regions is matched: the red channel value is mapped to the transparency of the cheek area, the green channel value is mapped to the saturation of the area between the eyebrows, and the blue channel value is mapped to the hue of the lip area, generating a basic static expression texture. If the expression parameters include blink frequency, the blink frequency is converted into the transparency change period, and the real-time transparency coefficient of each video frame is calculated using a sine function, the expression of which is...

[0130]

[0131] Where U(t) represents the real-time transparency coefficient; f represents the blink frequency parameter; and t represents the current frame timestamp.

[0132] A dynamic expression sequence is generated by overlaying a transparency coefficient onto a basic static facial expression texture, which is a facial expression.

[0133] The motion parameters are input into the motion planning operation. The current joint angle (the initial posture angle value of the digital human limb structure) is input into the inverse kinematics to obtain the current position coordinates of the end effector. The difference between the current position coordinates of the end effector and the target position coordinates is used as the position deviation vector. The position deviation vector is input into the Jacobian matrix mapping operation. A pseudo-inverse matrix solution operation is performed on the Jacobian matrix. The pseudo-inverse matrix and the position deviation vector are multiplied to generate the original joint angle adjustment amount. The ratio of the damping intensity percentage value to 100% is used as a normalized value between 0 and 1, which is the joint motion damping coefficient. The product of the joint motion damping coefficient and the original joint angle adjustment amount is the actual joint angle adjustment amount. The actual joint angle adjustment amount is superimposed on the current joint angle to update the joint state. The joint angles are iteratively adjusted until the end effector reaches the target position, and the joint angle sequence is output as the limb motion trajectory.

[0134] S5.3. Synchronous audio streams, facial expressions, and body movement trajectories are aligned with timestamps and fused with behaviors to generate interactive response behaviors.

[0135] It should be noted that the synchronous audio stream input timestamp marking operation marks discrete time points as audio timestamps; the facial expression input video frame serialization operation marks the rendering timestamp of each frame based on the video frame rate to generate facial expression timestamps; the limb motion trajectory input trajectory interpolation operation resamples the joint angle change sequence according to the video frame rate to generate joint angle timestamps; the audio timestamp, facial expression timestamp, and joint angle timestamp input time alignment operation aligns the time axis based on the video frame timestamp, the audio waveform is segmented according to the video frame timestamp, and the joint angle sequence is interpolated and matched according to the video frame timestamp; the aligned audio waveform, facial expression sequence, and joint angle sequence input behavior fusion operation, the audio waveform segments are superimposed to the corresponding time interval, the facial expression sequence is mapped to the digital human face mesh, and the joint angle sequence drives the rotation of the digital human skeleton to output interactive response behavior.

[0136] This scheme integrates user speech and visual data through a cross-modal attention mechanism to generate a semantic association topology graph. It uses a curvature-driven algorithm to dynamically optimize vertex distribution and calculate joint activity thresholds to construct a multi-dimensional association map. Combined with an information density gradient field, it divides high / low density regions. The high-density core region is simplified through graph structure to generate a semantic core encoding tensor that retains strong association features. The low-density edge region is compressed through edge semantic entropy to generate an edge feature encapsulation package. A dynamically optimized language model is constructed to achieve end-to-end decoding of the semantic core encoding tensor and the edge feature encapsulation package. Natural language responses are generated through parameter mapping and template filling. Cross-modal instruction conversion drives the digital human to generate synchronous audio streams, facial expressions, and limb movement trajectories, realizing a closed loop of semantic understanding, physical constraints, and behavior generation.

[0137] This embodiment also provides a computer device applicable to the AI ​​digital human interactive response method based on a large language model, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the AI ​​digital human interactive response method based on a large language model as proposed in the above embodiment.

[0138] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0139] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the interactive response method for AI digital humans based on a large language model as proposed in the above embodiments. 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 Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0140] In summary, this invention achieves accurate modeling of cross-modal relationships and significantly reduces semantic parsing errors by: projecting cross-modal fusion vectors into cross-modal feature vectors using feature space; constructing a semantic association topology graph; dynamically optimizing the vertex distribution of the semantic association topology graph; calculating joint activity thresholds using joint motion constraint tensors; embedding physical constraints into the response generation link; and realizing differential geometrical coordination between motion trajectories and acoustic-optical responses, thus completely eliminating millisecond-level delays generated by independent modules.

[0141] 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 it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An AI digital human interactive response method based on a large language model, characterized in that: include, The collected user voice and visual data are analyzed using natural language processing methods to generate cross-modal feature vectors. Then, cross-modal association parsing is performed on the cross-modal feature vectors to generate a semantic association topology graph. The vertex coordinates and joint activity thresholds of the semantic association topology graph are calculated using high digital human relevance. The specific steps are as follows: Vertex and edge weight data of the semantic association topology graph are used to generate joint motion constraint tensors through high digital human relevance, specifically including: Convert the data of each vertex in the semantic association topology graph into the three-dimensional spatial coordinates of the skeletal joints of the digital human; The edge weights are mapped to the range of motion limits between joints, and the joint motion constraint tensor is calculated using the ergonomic constraint formula. The vertex position is dynamically adjusted based on the joint motion constraint tensor, and the joint activity threshold corresponding to the edge weight is mapped to output the vertex coordinates and the joint activity threshold. Attention weight calibration is performed by inputting vertex coordinates and joint activity thresholds into a coordinate index library, and a multidimensional correlation map is output. Information density analysis is performed based on the multidimensional association graph to generate an information density gradient vector field, and a high-density core region and a low-density edge region are divided. A semantic core encoding tensor is generated in the high-density core region, and an edge feature encapsulation package is generated in the low-density edge region. A dynamically optimized language model is constructed, and the semantic core encoding tensor and edge feature encapsulation package are input to generate a natural language text response, which is then converted into a multimodal response instruction set. Control parameters are extracted from the multimodal response instruction set, and corresponding synchronous audio streams, facial expressions, and limb movement trajectories are generated through the control parameters. The synchronous audio streams, facial expressions, and limb movement trajectories are integrated to generate interactive response behavior.

2. The AI ​​digital human interactive response method based on a large language model as described in claim 1, characterized in that: The specific steps for generating the semantic association topology graph are as follows: User voice data is used to generate word-segmented text sequences through speech recognition, and visual data is used to extract visual feature vectors through convolution calculation. Word embedding is performed on the segmented text sequence to generate text vectors, which are then input together with the visual feature vectors into a sparse attention mechanism to generate cross-modal fusion vectors. Cross-modal fusion vectors generate cross-modal feature vectors through feature space projection; cross-modal feature vectors construct semantic association topology graphs through cross-modal association parsing.

3. The AI ​​digital human interactive response method based on a large language model as described in claim 1, characterized in that: The specific steps for outputting the multidimensional correlation graph are as follows. Spatial encoding of vertex coordinates, association of joint activity thresholds with corresponding vertex pairs to form a coordinate-joint mapping table, and injection of timestamps and information entropy labels to build a coordinate index library; The coordinates of neighboring vertices are retrieved by spatial indexing, the joint activity threshold is obtained by calling the coordinate-joint mapping table, and the association strength threshold function is constructed. The association transition probability is calculated. If the association transition probability is greater than the preset probability threshold, a multidimensional association map is generated.

4. The AI ​​digital human interactive response method based on a large language model as described in claim 3, characterized in that: The specific steps for dividing the high-density core area and the low-density edge area are as follows. Input the multidimensional correlation map into the information density analysis formula to generate an information density distribution map. Perform gradient operator operations on the information density distribution map to generate an information density gradient vector field. The information density gradient magnitude is calculated based on the information density gradient vector field and compared with a preset entropy threshold. If the information density gradient magnitude exceeds the entropy threshold, it is classified as a high-density core region; otherwise, it is classified as a low-density edge region.

5. The AI ​​digital human interactive response method based on a large language model as described in claim 4, characterized in that: The high-density core region generates a semantic core encoding tensor, and the low-density edge region generates an edge feature encapsulation packet. The specific steps are as follows. The high-density core region is simplified through graph structure and compressed by tensor decomposition to generate a semantic core encoding tensor; Low-density edge regions are processed by edge semantic entropy compression calculation and edge feature encoding to generate edge feature encapsulation packages.

6. The AI ​​digital human interactive response method based on a large language model as described in claim 5, characterized in that: The specific steps for constructing a dynamically optimized language model are as follows. A decoding and fusion layer is constructed based on the principle of topology reconstruction; a response optimization layer is constructed based on the principle of parameter mapping; and a language generation layer is constructed based on the principle of template filling. A dynamically optimized language model is constructed through a decoding fusion layer, a response optimization layer, and a language generation layer.

7. The AI ​​digital human interactive response method based on a large language model as described in claim 6, characterized in that: The conversion to a multimodal response instruction set involves the following specific steps. The semantic core encoding tensor and edge feature encapsulation package are input into the dynamically optimized language model, and natural language text response is generated through topological reconstruction, parameter mapping and template filling. Transform natural language text responses into multimodal response instruction sets through cross-modal instruction conversion.

8. The AI ​​digital human interactive response method based on a large language model as described in claim 1, characterized in that: The control parameters include voice parameters, facial expression parameters, and motion parameters.

9. The AI ​​digital human interactive response method based on a large language model as described in claim 8, characterized in that: The specific steps for generating the interactive response are as follows. Speech parameters are used to generate a synchronized audio stream through Mel spectrum generation and vocoder conversion; Facial expressions are generated by simulating facial muscles using facial expression parameters. Motion parameters are solved using inverse kinematics to generate limb motion trajectories; By aligning synchronized audio streams, facial expressions, and body movement trajectories with timestamps and fusing them with behaviors, interactive response behaviors are generated.

Images