Traditional Chinese medicine intelligent diagnosis system based on multi-source data fusion and ai technology

Abstract

The application relates to the field of traditional Chinese medicine diagnosis and discloses a traditional Chinese medicine intelligent diagnosis system based on multi-source data fusion and AI technology, which comprises a tongue appearance collection and classification unit, is used for collecting tongue appearance data of a patient through an imaging device, performing feature extraction on the tongue appearance data, and obtaining a tongue appearance feature data set; performing tongue appearance pathological mode classification on the tongue appearance feature data set based on a deep convolutional neural network to obtain tongue diagnosis recognition result data; and a facial diagnosis feature correlation analysis unit, which is used for inputting the tongue diagnosis recognition result data into a multi-modal data fusion engine, performing facial diagnosis pathological feature correlation analysis based on a graph attention network, and obtaining facial diagnosis recognition result data; the application fuses three traditional diagnosis methods of tongue diagnosis, facial diagnosis and pulse diagnosis, introduces various data collection modes such as three-dimensional imaging, thermal imaging, micro-expression recognition, spectral information and pulse appearance harmonic analysis, and realizes deep semantic understanding by combining advanced models such as a graph neural network, a U-Net and a Bi-LSTM.

Description

Technical Field

[0001] This invention relates to the field of TCM diagnostic technology, specifically a TCM intelligent diagnostic system based on multi-source data fusion and AI technology. Background Technology

[0002] The four diagnostic methods of traditional Chinese medicine—inspection, auscultation and olfaction, inquiry, and palpation—are its core diagnostic tools, with tongue diagnosis, facial diagnosis, and pulse diagnosis being the three most objective and visually appealing. Traditional Chinese medicine diagnosis relies on the doctor's experience, which leads to problems such as strong subjectivity, difficulty in standardization, and low efficiency.

[0003] In recent years, with the development of medical image processing, sensing technology, and artificial intelligence algorithms, researchers have attempted to combine computer vision, signal processing, and traditional Chinese medicine (TCM) diagnosis to achieve automatic recognition and diagnostic reasoning of tongue, facial, and pulse images. However, existing systems are mostly limited to single-modal recognition, such as relying solely on tongue images or single pulse signals, lacking a fusion understanding of multi-source TCM data, and failing to fully reflect the TCM concepts of "holistic view" and "treatment based on syndrome differentiation."

[0004] Furthermore, existing methods still have significant shortcomings in terms of diagnostic accuracy, generalization ability, and interpretability when faced with the heterogeneity, unstructured nature, and individual differences among samples in multimodal data. Therefore, there is an urgent need to construct an intelligent diagnostic system that integrates multi-source information from tongue diagnosis, facial diagnosis, and pulse diagnosis, and possesses deep semantic modeling capabilities, in order to improve the scientific and intelligent level of TCM diagnosis. Summary of the Invention

[0005] The purpose of this invention is to provide a TCM intelligent diagnostic system based on multi-source data fusion and AI technology, in order to solve the problems that the existing systems are mostly limited to single-modal recognition, based only on tongue images or single pulse signals, lack the fusion and understanding of multi-source TCM data, and are difficult to fully restore the TCM concepts of "holistic view" and "differentiation of syndromes and treatment". They still have significant deficiencies in terms of diagnostic accuracy, generalization ability and interpretability.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a TCM intelligent diagnostic system based on multi-source data fusion and AI technology, comprising:

[0007] The tongue image acquisition and classification unit is used to acquire patients' tongue image data through imaging equipment, extract features from the tongue image data to obtain a tongue image feature dataset; and classify the tongue image pathological patterns of the tongue image feature dataset based on a deep convolutional neural network to obtain tongue diagnosis identification results data.

[0008] The facial diagnosis feature association analysis unit is used to input the tongue diagnosis identification result data into the multimodal data fusion engine, collect the patient's facial data, extract features from the facial data to obtain a facial diagnosis feature dataset; combine the tongue diagnosis identification result data to dynamically assign weights to the facial diagnosis feature dataset, and perform facial diagnosis pathological feature association analysis based on a graph attention network to obtain facial diagnosis identification result data.

[0009] The pulse diagnosis feature identification and processing unit is used to input the facial diagnosis identification result data into the pulse diagnosis and consultation joint analysis module, collect the patient's pulse feature data, perform time-frequency joint analysis and waveform feature clustering on the pulse signal to obtain the pulse diagnosis feature dataset; combine the facial diagnosis identification result data with the pulse diagnosis feature dataset to perform syndrome pulse mapping matching, and perform pulse diagnosis pathological feature identification based on a bidirectional long short-term memory network to obtain the pulse diagnosis identification result data.

[0010] The three-diagnosis fusion decision push unit is used to input the tongue diagnosis identification results data, face diagnosis identification results data, and pulse diagnosis identification results data into the multi-source data fusion decision module, construct the three-diagnosis information association matrix and perform feature weight allocation; based on graph neural network, it performs four-diagnosis information fusion reasoning, generates TCM syndrome diagnosis result data and pushes it to the terminal device to perform TCM intelligent diagnosis and auxiliary decision-making.

[0011] Furthermore, the facial data features include micro-expression recognition and blood circulation distribution, and the patient's facial data includes facial three-dimensional point cloud data and thermal imaging data.

[0012] Furthermore, tongue image data of the patient is acquired through imaging equipment to obtain a tongue image feature dataset, including:

[0013] The patient's tongue image data was acquired using an imaging device. The tongue image data included the reflectance spectrum information of the tongue surface and the transmission spectrum information of the subcutaneous tissue. The patient's tongue image data was then processed by a dynamic exposure compensation algorithm to equalize the brightness, resulting in the original tongue image data.

[0014] The original tongue image data is input into the three-dimensional tongue modeling module, and the structured light projection technology is used to obtain the micron-level morphological data of the tongue surface. The dynamic deformation correction of the tongue is achieved by combining the elastic deformation registration algorithm. The three-dimensional geometric model of the tongue is constructed by multi-view image stitching technology to obtain the three-dimensional tongue image data.

[0015] Three-dimensional tongue image data is input into a multi-dimensional feature extraction engine, and an adversarial generative network is used to enhance the tongue image data to generate a simulated tongue image dataset. Based on the improved U-Net semantic segmentation network, the tongue image is divided into regions to obtain a structured tongue image feature dataset.

[0016] The structured tongue image feature dataset is input into the feature optimization and fusion module. The self-attention mechanism is used to assign weights to features of different dimensions. Cross-modal feature association analysis is achieved through graph convolutional networks. A contrastive learning strategy is used to eliminate interference from individual differences, resulting in a standardized tongue image feature vector set.

[0017] Furthermore, based on a deep convolutional neural network, the tongue image feature dataset is classified into tongue pathological patterns to obtain tongue diagnosis identification results data, including:

[0018] The standardized tongue image feature vector set is input into the multi-scale feature perception module to extract local and global features of tongue image texture, color and edge to obtain a multi-scale feature map of the tongue image; the channel attention mechanism is used to dynamically assign weights to the multi-scale feature map of the tongue image to generate an enhanced feature map.

[0019] The enhanced feature map is input into the dynamic path selection network; the contribution of each path to the classification of the current tongue image sample is evaluated in real time through a reinforcement learning agent, resulting in customized feature representation data.

[0020] Customized feature representation data is input into the cross-modal association learning module, and semantic association between tongue image features and pathological syndromes is constructed through graph neural network. A contrastive learning strategy is adopted to maximize the feature similarity of positive sample pairs, while a memory bank mechanism is used to store historical sample features to enhance the model's ability to identify rare pathological patterns, resulting in semantically enhanced tongue image feature embedding data.

[0021] By embedding semantically enhanced tongue image features into the data input uncertainty-aware classification head, tongue diagnosis identification results are obtained.

[0022] Furthermore, the tongue diagnosis results are input into a multimodal data fusion engine to obtain a facial diagnosis feature dataset, including:

[0023] The tongue diagnosis results data are input into the prior knowledge injection module of the multimodal data fusion engine to obtain the tongue diagnosis prior heatmap;

[0024] A spectral 3D facial scanner was used to acquire dynamic micro-expression video streams of patients’ facial structured light data and texture reflectance spectral data to generate multi-dimensional raw facial data. The tongue diagnosis prior heat map was superimposed with the facial scan data, and the network was guided to focus on highly correlated areas through a spatial attention mechanism to generate facial input data.

[0025] Facial input data is input into a dynamic feature extraction network, which includes a static branch structure and a dynamic branch structure. The dynamic feature extraction network performs semantic alignment and fusion through a cross-modal interaction module to obtain joint feature representation data for tongue diagnosis and facial diagnosis.

[0026] The combined feature representation data of tongue diagnosis and facial diagnosis is input into the feature decoupling and reconstruction module, which includes:

[0027] Feature decoupling is achieved by separating pathological components and individual difference noise in facial features through a categorical conditional variational autoencoder.

[0028] The reconstruction module dynamically adjusts the feature reconstruction intensity by combining the pathological severity score in the tongue diagnosis results, and generates a standardized facial diagnosis feature vector set.

[0029] Furthermore, by dynamically assigning weights to the facial diagnosis feature dataset in conjunction with the tongue diagnosis identification data, facial diagnosis identification data is obtained, including:

[0030] The tongue diagnosis results are input into the dynamic weight generation module. A pathological association knowledge graph is constructed through the TCM tongue surface correspondence theory, and tongue diagnosis keywords are mapped to facial diagnosis feature weight parameters. At the same time, a reinforcement learning framework is introduced, with the maximization of mutual information between facial diagnosis features and tongue diagnosis results as the reward function, to obtain tongue diagnosis weight allocation strategy data.

[0031] A facial imaging system is used to collect multimodal facial data of patients to generate a raw facial diagnosis dataset. The tongue diagnosis weight parameters are spatially superimposed with the facial data, and the network is guided to focus on high-weight areas through an attention mechanism to generate tongue diagnosis-enhanced facial diagnosis input data.

[0032] The facial diagnosis input data enhanced by tongue diagnosis is input into a cross-modal feature extraction network to obtain dynamic facial functional features;

[0033] The cross-modal feature extraction network integrates features through a cross-modal interaction module and adjusts feature weights based on the pathological degree of tongue diagnosis to obtain secondary weighted facial diagnosis feature data.

[0034] The weighted facial diagnosis feature data is input into a dynamic graph attention network to obtain the patient's facial and tongue diagnosis heterogeneity graph:

[0035] A multi-head attention mechanism is used to propagate pathological information on the heteromorphic map of the patient's face and tongue diagnosis, focusing on strengthening the information flow between highly correlated nodes of face and tongue diagnosis, and generating pathological correlation feature data of face diagnosis.

[0036] The pathological correlation feature data is input into the decoupled and reconstructed diagnostic module to generate standardized facial diagnosis and identification results data.

[0037] Furthermore, the facial diagnosis results are input into the pulse diagnosis and medical history joint analysis module to obtain the pulse diagnosis feature dataset, including:

[0038] The facial diagnosis results are input into the pulse diagnosis and consultation joint analysis module. A knowledge graph of syndrome and pulse correlation is constructed through the TCM theory of facial and pulse correspondence. Facial diagnosis keywords are mapped to pulse diagnosis feature weight parameters. A reinforcement learning framework is introduced, with the maximization of mutual information between pulse features and facial diagnosis results as the reward function, to generate initial weight allocation strategy data.

[0039] A wearable pulse sensor array is used, combined with initial weight allocation strategy data, to dynamically adjust the sampling strategy and generate raw pulse signal data;

[0040] The original pulse signal data is input into the time-frequency analysis module, and the pulse harmonic components are extracted using adaptive wavelet transform. The characteristic frequency bands are then highlighted by weighting the facial diagnosis weight parameters. Subsequently, the pulse signal is decomposed into multiple intrinsic mode functions using variational mode decomposition, and a pulse diagnosis feature dataset is generated by using a spectral clustering algorithm guided by facial diagnosis syndrome types.

[0041] Furthermore, by combining the facial diagnosis results with the pulse diagnosis feature dataset, syndrome pulse mapping and matching are performed to obtain pulse diagnosis results data, including:

[0042] The facial diagnosis results are input into a dynamic weight allocation engine to extract the correlation weight matrix between facial meridian energy distribution features and tongue diagnosis syndrome labels; based on this weight matrix, the pulse diagnosis feature dataset is recalibrated to generate a pulse energy distribution map.

[0043] Based on the syndrome types in the facial diagnosis results data, a search and matching is performed in the pre-built syndrome pulse association knowledge base to obtain typical pulse feature patterns. The typical pulse feature patterns are then compared with the initial pulse diagnosis feature dataset to obtain a subset of pulse diagnosis features.

[0044] A subset of pulse diagnosis features is input into a bidirectional long short-term memory network model trained with a large amount of pulse diagnosis data. Through the forward and backward propagation process of the bidirectional long short-term memory network, deep learning and feature mining are performed on the subset of pulse diagnosis features to obtain the pulse diagnosis feature vector.

[0045] The similarity calculation and matching analysis are performed between the pulse diagnosis feature vector and the pre-set standard template of pulse feature corresponding to different syndromes. The matching results are then comprehensively evaluated and adjusted by combining the syndrome weight information in the face diagnosis identification results data to obtain the pulse diagnosis identification results data.

[0046] Furthermore, the results of tongue diagnosis, facial diagnosis, and pulse diagnosis are input into the multi-source data fusion decision module to generate TCM syndrome diagnosis results, including:

[0047] The results of tongue diagnosis, face diagnosis, and pulse diagnosis are input into the multi-source data fusion decision module. The pathological keywords of each diagnosis result are analyzed through the TCM three-diagnosis association knowledge graph to generate a structured feature vector set.

[0048] Based on the structured feature vector set, a multi-dimensional correlation analysis algorithm is used to generate the correlation matrix of the three diagnostic methods information:

[0049] The three-diagnosis information association matrix includes horizontal association, vertical reinforcement and conflict resolution. The three-diagnosis information association matrix is ​​adjusted by posterior probability through a Bayesian network to dynamically optimize the weight allocation and obtain a three-dimensional weight matrix.

[0050] The three-dimensional weight matrix and the structured feature vector set are input into the graph neural network fusion engine to construct the patient's four diagnostic information graph: the four diagnostic information graph includes node construction, edge dynamic evolution and integration, and attention propagation;

[0051] Traditional Chinese medicine syndrome diagnosis data is generated through a graph classification layer. The TCM syndrome diagnosis data includes syndrome type, confidence level, and cross-modal evidence chain.

[0052] Furthermore, in the multi-source data fusion decision-making module, a comprehensive contribution calculation formula is introduced, as follows:

[0053] Let the tongue diagnosis identification results data be The facial diagnosis and identification results data are The pulse diagnosis results data are as follows ,in , , These represent the number of feature dimensions in the identification results data for tongue diagnosis, facial diagnosis, and pulse diagnosis, respectively.

[0054] Define the characteristics of tongue diagnosis The contribution to the final diagnostic result is Facial features The contribution to the final diagnostic result is Pulse diagnosis characteristics The contribution to the final diagnostic result is ,in ;

[0055] The contribution of each feature is calculated based on the correlation between the feature and the final syndrome diagnosis result, as well as the importance of the feature in its respective diagnostic module, and is calculated in the following way:

[0056] For tongue diagnosis characteristics Its contribution The calculation formula is:

[0057]

[0058] in, Indicating the characteristics of tongue diagnosis Data related to the final TCM syndrome diagnosis covariance, and These respectively represent the characteristics of tongue diagnosis. The variance of the final TCM syndrome diagnosis data D; Indicating the characteristics of tongue diagnosis Importance score in the classification of tongue diagnosis pathological patterns (can be obtained through feature selection algorithms or model interpretability methods); and It is the adjustment coefficient, and + =1, used to balance the impact of the correlation between the feature and the diagnostic result and the importance of the feature itself on the contribution;

[0059] Similarly, for facial diagnostic features Its contribution The calculation formula is:

[0060]

[0061] in Indicates facial features Data related to the final TCM syndrome diagnosis covariance, Indicates facial features The variance; Indicates facial features Importance score in the correlation analysis of pathological features during face diagnosis; and It is the adjustment coefficient, and + =1.

[0062] For pulse diagnosis characteristics Its contribution The calculation formula is:

[0063]

[0064] in, Indicating pulse diagnosis characteristics Data related to the final TCM syndrome diagnosis covariance, Indicating pulse diagnosis characteristics The variance; Indicating pulse diagnosis characteristics The importance score in the process of identifying pathological features in pulse diagnosis; and It is the adjustment coefficient, and + =1; Finally, by combining the characteristic contributions of tongue diagnosis, facial diagnosis, and pulse diagnosis, the comprehensive contribution weight of each diagnostic module to the final TCM syndrome diagnosis result is obtained. :

[0065]

[0066]

[0067]

[0068] And on Perform normalization to make it meet the requirements. + + =1, the normalized weight is used to more reasonably integrate the identification results of tongue diagnosis, face diagnosis and pulse diagnosis in the multi-source data fusion decision module, so as to generate more accurate TCM syndrome diagnosis results.

[0069] Compared with the prior art, the beneficial effects of the present invention are:

[0070] (1) This invention integrates three traditional diagnostic methods—tongue diagnosis, face diagnosis, and pulse diagnosis—to construct a multimodal information complementarity mechanism, thereby achieving comprehensive perception and identification of the patient's physical condition and pathological state.

[0071] (2) This invention introduces multiple data acquisition methods such as three-dimensional imaging, thermal imaging, micro-expression recognition, spectral information and pulse harmonic analysis, and combines advanced models such as graph neural networks, U-Net, and Bi-LSTM to achieve deep semantic understanding.

[0072] (3) This invention enhances the ability to identify abnormal samples and pathological samples with few samples by using techniques such as adversarial generative networks, variational mode decomposition, and contrastive learning, thereby improving the model’s ability to adapt to individual differences.

[0073] (4) This invention embeds TCM theoretical knowledge such as tongue and pulse correspondence, meridian differentiation, and syndrome deduction into the system in the form of a structured map to achieve knowledge-driven data fusion and intelligent decision-making.

[0074] (5) This system generates diagnostic results and outputs confidence level and cross-modal evidence chain at the same time, which has good interpretability and traceability, and enhances the trust of clinicians. Attached Figure Description

[0075] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0076] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0077] Please see Figure 1 This invention provides a technical solution: a TCM intelligent diagnostic system based on multi-source data fusion and AI technology, comprising:

[0078] The tongue image acquisition and classification unit is used to acquire patients' tongue image data through imaging equipment, extract features from the tongue image data to obtain a tongue image feature dataset; and classify the tongue image pathological patterns of the tongue image feature dataset based on a deep convolutional neural network to obtain tongue diagnosis identification results data.

[0079] Tongue diagnosis, in Traditional Chinese Medicine (TCM), involves doctors observing the color, shape, and coating of a patient's tongue to determine their constitution and pathological state. Tongue imagery is a comprehensive image of the tongue's appearance. Imaging equipment is used to collect image data of the tongue, such as high-definition cameras or dedicated tongue image acquisition devices. Feature extraction involves extracting representative data from the image that can be used for machine learning modeling, such as tongue coating area, color distribution, and tongue outline. The tongue image feature dataset contains a large amount of processed and labeled tongue image feature sample data for training and testing models. Deep convolutional neural networks (CNNs) are highly effective deep learning models for image processing, excelling at automatically extracting spatial features from images and classifying them. Tongue pathological pattern classification uses CNN models to categorize tongue images into different TCM pathological types (such as pale tongue, thick and greasy tongue coating, etc.) for diagnostic purposes.

[0080] The facial diagnosis feature association analysis unit is used to input the tongue diagnosis identification results data into the multimodal data fusion engine, collect the patient's facial data, extract features from the facial data, and obtain the facial diagnosis feature dataset; combine the tongue diagnosis identification results data to dynamically assign weights to the facial diagnosis feature dataset, and perform facial diagnosis pathological feature association analysis based on graph attention network to obtain the facial diagnosis identification results data.

[0081] The system integrates information from different sensory channels (such as images, sounds, and pulse) and unifies the analysis and modeling. Facial data acquisition uses cameras or infrared devices to capture images of patients' faces to identify facial features such as color, skin texture, wrinkles, and shape. The facial feature dataset contains key diagnostic features of the patient's face, such as a sallow complexion, flushed face, and dark circles, for the AI ​​model to learn and judge. Dynamic weight allocation assigns different analytical weights to data from different sources based on their importance or credibility, making multimodal fusion more accurate. The graph attention network is a graph neural network model that can dynamically adjust the contribution of each node (such as features) to the overall result when analyzing graph-structured data.

[0082] The pulse diagnosis feature identification and processing unit is used to input the facial diagnosis identification result data into the pulse diagnosis and consultation joint analysis module, collect the patient's pulse feature data, perform time-frequency joint analysis and waveform feature clustering on the pulse signal to obtain the pulse diagnosis feature dataset; combine the facial diagnosis identification result data with the pulse diagnosis feature dataset to perform syndrome pulse mapping matching, and perform pulse diagnosis pathological feature identification based on a bidirectional long short-term memory network to obtain the pulse diagnosis identification result data.

[0083] Among them, pulse feature data is obtained from wrist artery pulse signals acquired by pulse diagnosis instruments, such as pulse frequency, amplitude, and rhythm; time-frequency joint analysis analyzes pulse signals in both time and frequency domains to extract more comprehensive features (such as waveform shape changes over time); waveform feature clustering groups pulse signals with similar waveform features into one category to extract common pulse patterns; syndrome pulse mapping establishes a correspondence between TCM syndromes (such as qi deficiency, blood stasis, damp heat, etc.) and pulse data to assist in diagnosis; bidirectional long short-term memory network is a recurrent neural network that can consider past and future pulse feature changes simultaneously in time series modeling to improve identification accuracy;

[0084] The three-diagnosis fusion decision push unit is used to input the tongue diagnosis identification results data, face diagnosis identification results data, and pulse diagnosis identification results data into the multi-source data fusion decision module, construct the three-diagnosis information association matrix and perform feature weight allocation; based on graph neural network, it performs four-diagnosis information fusion reasoning, generates TCM syndrome diagnosis result data and pushes it to the terminal device to perform TCM intelligent diagnosis and auxiliary decision-making.

[0085] The multi-source data fusion decision module integrates data from three sources—tongue diagnosis, facial diagnosis, and pulse diagnosis—and uses algorithms to arrive at a unified diagnostic result. The three-diagnosis information association matrix constructs a feature relationship network between the three diagnostic methods, quantifying their correlation. Feature weight allocation automatically adjusts the influence of different features on the final diagnostic result based on their importance. The graph neural network is a deep learning model suitable for analyzing graph-structured data, capable of handling complex relationships between nodes, and is used for information fusion and reasoning in the four diagnostic methods. The TCM syndrome diagnosis result is the final output as a TCM syndrome judgment result, such as "liver fire excess," "qi stagnation and blood stasis," and "spleen and stomach weakness," for doctors' reference. The diagnostic results are pushed to the terminal device and transmitted to the doctor's operating interface or the patient's APP, supporting clinical decision-making or health advice.

[0086] It should be noted that during operation, the system uses three typical TCM diagnostic methods—tongue diagnosis, facial diagnosis, and pulse diagnosis—to make a collaborative judgment, avoiding bias caused by a single diagnosis and achieving more comprehensive and accurate TCM syndrome identification. A deep learning model is introduced to automatically extract and analyze key features, reducing subjective human interference and improving diagnostic efficiency and standardization. A multimodal fusion engine and graph neural network are employed to fully utilize the inherent connections between various diagnostic information, enhancing overall reasoning ability and decision-making credibility. The system is divided into four main units: data collection, analysis, fusion, and push. The modular design facilitates subsequent functional expansion, model updates, and integration with other TCM information systems. The diagnostic results are ultimately pushed to terminal devices (such as doctor workstations and patient apps) for real-time feedback, facilitating doctor-assisted decision-making or remote health management.

[0087] In one embodiment, facial data features include micro-expression recognition and blood flow distribution, and patient facial data includes facial three-dimensional point cloud data and thermal imaging data.

[0088] In one embodiment, tongue image data of a patient is acquired using an imaging device to obtain a tongue image feature dataset, including:

[0089] The patient's tongue image data was acquired using an imaging device. The patient's tongue image data included the reflectance spectrum information of the tongue surface and the transmission spectrum information of the subcutaneous tissue. The patient's tongue image data was then processed for brightness equalization using a dynamic exposure compensation algorithm to obtain the original tongue image data.

[0090] The original tongue image data is input into the 3D tongue modeling module, and the structured light projection technology is used to obtain the micron-level morphological data of the tongue surface. The dynamic deformation correction of the tongue is achieved by combining the elastic deformation registration algorithm. The 3D geometric model of the tongue is constructed by multi-view image stitching technology to obtain the 3D tongue image data.

[0091] Three-dimensional tongue image data is input into a multi-dimensional feature extraction engine, and an adversarial generative network is used to enhance the tongue image data to generate a simulated tongue image dataset. Based on the improved U-Net semantic segmentation network, the tongue image is divided into regions to obtain a structured tongue image feature dataset.

[0092] The structured tongue image feature dataset is input into the feature optimization and fusion module. The self-attention mechanism is used to assign weights to features of different dimensions. Cross-modal feature association analysis is achieved through graph convolutional networks. A contrastive learning strategy is used to eliminate interference from individual differences, resulting in a standardized tongue image feature vector set.

[0093] This design achieves high-precision modeling of the tongue image by fusing reflectance and transmission spectra, combined with 3D structured light and deformation correction. Multi-view stitching solves the occlusion problem and enhances model integrity. The introduction of GAN data augmentation and U-Net semantic segmentation ensures rich feature structure and accurate partitioning. Contrastive learning and graph convolution further enhance feature universality and semantic expression capabilities, making tongue image information acquisition more comprehensive and feature expression more standardized, effectively improving the accuracy of subsequent tongue diagnosis classification.

[0094] In one embodiment, a deep convolutional neural network is used to classify tongue pathological patterns in a tongue image feature dataset to obtain tongue diagnosis identification results data, including:

[0095] The standardized tongue image feature vector set is input into the multi-scale feature perception module to extract local and global features of tongue image texture, color and edge to obtain a multi-scale feature map of the tongue image; the channel attention mechanism is used to dynamically assign weights to the multi-scale feature map of the tongue image to generate an enhanced feature map.

[0096] The enhanced feature map is input into the dynamic path selection network; the contribution of each path to the classification of the current tongue image sample is evaluated in real time through a reinforcement learning agent, resulting in customized feature representation data.

[0097] Customized feature representation data is input into the cross-modal association learning module, and semantic association between tongue image features and pathological syndromes is constructed through graph neural network. A contrastive learning strategy is adopted to maximize the feature similarity of positive sample pairs, while a memory bank mechanism is used to store historical sample features to enhance the model's ability to identify rare pathological patterns, resulting in semantically enhanced tongue image feature embedding data.

[0098] By embedding semantically enhanced tongue image features into the data input uncertainty-aware classification head, tongue diagnosis identification results are obtained.

[0099] This design extracts rich texture and structural features through a multi-scale perception module, strengthens key patterns through channel attention, introduces personalized learning strategies through a dynamic path selection network to enhance the model's adaptability to complex tongue appearances, establishes deep semantic relationships between tongue appearances and pathological labels through a graph neural network, and enhances the recognition of rare signs through a memory mechanism. The overall design improves the flexibility, generalization, and diagnostic confidence of tongue appearance pathology recognition.

[0100] In one embodiment, the tongue diagnosis result data is input into a multimodal data fusion engine to obtain a facial diagnosis feature dataset, including:

[0101] The tongue diagnosis results data are input into the prior knowledge injection module of the multimodal data fusion engine to obtain the tongue diagnosis prior heatmap;

[0102] A spectral 3D facial scanner was used to acquire dynamic micro-expression video streams of patients’ facial structured light data and texture reflectance spectral data to generate multi-dimensional raw facial data. The tongue diagnosis prior heat map was superimposed with the facial scan data, and the network was guided to focus on highly correlated areas through a spatial attention mechanism to generate facial input data.

[0103] Facial input data is fed into a dynamic feature extraction network, which includes static and dynamic branch structures. The dynamic feature extraction network performs semantic alignment and fusion through a cross-modal interaction module to obtain joint feature representation data for tongue and facial diagnosis.

[0104] The combined feature representation data of tongue diagnosis and facial diagnosis is input into the feature decoupling and reconstruction module. The feature decoupling and reconstruction module includes:

[0105] Feature decoupling is achieved by separating pathological components and individual difference noise in facial features through a categorical conditional variational autoencoder.

[0106] The reconstruction module dynamically adjusts the feature reconstruction intensity by combining the pathological severity score in the tongue diagnosis results, and generates a standardized facial diagnosis feature vector set.

[0107] This design uses tongue diagnosis heatmaps as prior knowledge to guide facial feature extraction, effectively focusing on areas related to TCM pathology. It adopts a static and dynamic dual-branch structure, integrating the three-dimensional structure of the face with micro-expression changes to enhance the dynamic sensitivity of facial diagnosis. Through feature decoupling and reconstruction, it distinguishes between pathological and individual differences, making facial diagnosis features more standardized and generalizable. The overall design achieves tongue-face complementarity, enhancing the TCM explanatory power of facial diagnosis features.

[0108] In one embodiment, dynamic weight allocation is performed on the facial diagnosis feature dataset in conjunction with the tongue diagnosis identification result data to obtain facial diagnosis identification result data, including:

[0109] The tongue diagnosis results are input into the dynamic weight generation module. A pathological association knowledge graph is constructed through the TCM tongue surface correspondence theory, and tongue diagnosis keywords are mapped to facial diagnosis feature weight parameters. At the same time, a reinforcement learning framework is introduced, with the maximization of mutual information between facial diagnosis features and tongue diagnosis results as the reward function, to obtain tongue diagnosis weight allocation strategy data.

[0110] A facial imaging system was used to collect multimodal facial data from patients to generate a raw facial diagnosis dataset. Tongue diagnosis weight parameters were spatially superimposed on the facial data, and an attention mechanism was used to guide the network to focus on high-weight regions to generate tongue diagnosis-enhanced facial diagnosis input data.

[0111] The facial diagnosis input data enhanced by tongue diagnosis is input into a cross-modal feature extraction network to obtain dynamic facial functional features;

[0112] The cross-modal feature extraction network integrates features through a cross-modal interaction module and adjusts feature weights based on the pathological degree of tongue diagnosis to obtain secondary weighted facial diagnosis feature data.

[0113] The weighted facial diagnosis feature data is input into a dynamic graph attention network to obtain the patient's facial and tongue diagnosis heterogeneity graph:

[0114] A multi-head attention mechanism is used to propagate pathological information on the heteromorphic map of the patient's face and tongue diagnosis, focusing on strengthening the information flow between highly correlated nodes of face and tongue diagnosis, and generating pathological correlation feature data of face diagnosis.

[0115] The pathological correlation feature data is input into the decoupled and reconstructed diagnostic module to generate standardized facial diagnosis and identification results data.

[0116] This design utilizes tongue diagnosis results to construct a pathological knowledge graph, guiding the weight allocation of facial diagnosis features and reflecting the TCM theory of "correspondence between tongue and face." It introduces reinforcement learning to optimize mutual information and enhance the complementarity of cross-modal features. The dynamic graph attention network provides fine control over the information flow in the heterogeneous graphs of facial and tongue diagnosis, improving the accuracy of associated feature recognition. This design strengthens the role of facial diagnosis in overall diagnosis and improves the relevance and robustness of identification.

[0117] In one embodiment, the facial diagnosis identification results data are input into the pulse diagnosis and questioning joint analysis module to obtain a pulse diagnosis feature dataset, including:

[0118] The facial diagnosis results are input into the pulse diagnosis and consultation joint analysis module. A knowledge graph of syndrome and pulse correlation is constructed through the TCM theory of facial and pulse correspondence. Facial diagnosis keywords are mapped to pulse diagnosis feature weight parameters. A reinforcement learning framework is introduced, with the maximization of mutual information between pulse features and facial diagnosis results as the reward function, to generate initial weight allocation strategy data.

[0119] A wearable pulse sensor array is used, combined with initial weight allocation strategy data, to dynamically adjust the sampling strategy and generate raw pulse signal data;

[0120] The original pulse signal data is input into the time-frequency analysis module, and the pulse harmonic components are extracted using adaptive wavelet transform. The characteristic frequency bands are then highlighted by weighting the facial diagnosis weight parameters. Subsequently, the pulse signal is decomposed into multiple intrinsic mode functions using variational mode decomposition, and a pulse diagnosis feature dataset is generated by using a spectral clustering algorithm guided by facial diagnosis syndrome types.

[0121] This design constructs a knowledge graph based on the correspondence between the face and pulse, guiding the pulse sampling strategy and improving the targeting of pulse data. It utilizes time-frequency analysis and variational mode decomposition to extract multi-scale waveform features, and combines them with facial diagnosis information for supervised clustering to optimize feature expression. This design makes full use of prior facial diagnosis information, making pulse sampling more diagnostically targeted and improving the accuracy of structural recognition and classification of complex pulse signals.

[0122] In one embodiment, pulse diagnosis feature dataset is mapped and matched with facial diagnosis identification data to obtain pulse diagnosis identification result data, including:

[0123] The facial diagnosis results are input into a dynamic weight allocation engine to extract the correlation weight matrix between facial meridian energy distribution features and tongue diagnosis syndrome labels; based on this weight matrix, the pulse diagnosis feature dataset is recalibrated to generate a pulse energy distribution map.

[0124] Based on the syndrome types in the facial diagnosis results data, a search and matching is performed in the pre-built syndrome pulse association knowledge base to obtain typical pulse feature patterns. The typical pulse feature patterns are then compared with the initial pulse diagnosis feature dataset to obtain a subset of pulse diagnosis features.

[0125] A subset of pulse diagnosis features is input into a bidirectional long short-term memory network model trained with a large amount of pulse diagnosis data. Through the forward and backward propagation process of the bidirectional long short-term memory network, deep learning and feature mining are performed on the subset of pulse diagnosis features to obtain the pulse diagnosis feature vector.

[0126] The similarity calculation and matching analysis are performed between the pulse diagnosis feature vector and the pre-set standard template of pulse feature corresponding to different syndromes. The matching results are then comprehensively evaluated and adjusted by combining the syndrome weight information in the face diagnosis identification results data to obtain the pulse diagnosis identification results data.

[0127] This design transforms the results of facial diagnosis into meridian energy distribution and syndrome labels, recalibrates pulse diagnosis features, and then inputs them into Bi-LSTM for deep learning, making the pulse diagnosis results more consistent with the patient's overall performance. Combining the syndrome database for matching and similarity analysis enhances the interpretability of the model. This design significantly improves the accuracy of pulse diagnosis and enhances the model's recognition performance in complex syndromes by incorporating context.

[0128] In one embodiment, tongue diagnosis results, facial diagnosis results, and pulse diagnosis results are input into a multi-source data fusion decision module to generate TCM syndrome diagnosis results, including:

[0129] The results of tongue diagnosis, face diagnosis, and pulse diagnosis are input into the multi-source data fusion decision module. The pathological keywords of each diagnosis result are analyzed through the TCM three-diagnosis association knowledge graph to generate a structured feature vector set.

[0130] Based on the structured feature vector set, a multi-dimensional correlation analysis algorithm is used to generate the correlation matrix of the three diagnostic methods information:

[0131] The association matrix of the three diagnostic information includes horizontal association, vertical reinforcement and conflict resolution. By adjusting the posterior probability of the association matrix of the three diagnostic information through a Bayesian network, the weight allocation is dynamically optimized to obtain a three-dimensional weight matrix.

[0132] The three-dimensional weight matrix and the structured feature vector set are input into the graph neural network fusion engine to construct the patient's four diagnostic information graph: the four diagnostic information graph includes node construction, edge dynamic evolution and integration, and attention propagation;

[0133] Traditional Chinese medicine syndrome diagnosis data is generated through graph classification layers. The TCM syndrome diagnosis data includes syndrome type, confidence level, and cross-modal evidence chain.

[0134] This design integrates the identification results of tongue diagnosis, facial diagnosis, and pulse diagnosis, constructs structured feature vectors through knowledge graphs, and optimizes information association weights using Bayesian methods. Graph neural networks complete the information fusion, conflict reconciliation, and reasoning among the three diagnostic data, ultimately generating syndrome types and confidence levels. This design improves the systematicness and automation level of the four diagnostic methods and realizes intelligent comprehensive identification of multi-source information in traditional Chinese medicine.

[0135] In one embodiment, a comprehensive contribution calculation formula is introduced in the multi-source data fusion decision module, as follows: Let the tongue diagnosis identification result data be... The facial diagnosis and identification results data are The pulse diagnosis results data are as follows ,in , , These represent the number of feature dimensions in the identification results data for tongue diagnosis, facial diagnosis, and pulse diagnosis, respectively.

[0136] Define the characteristics of tongue diagnosis The contribution to the final diagnostic result is Facial features The contribution to the final diagnostic result is Pulse diagnosis characteristics The contribution to the final diagnostic result is ,in ;

[0137] The contribution of each feature is calculated based on the correlation between the feature and the final syndrome diagnosis result, as well as the importance of the feature in its respective diagnostic module, and is calculated in the following way:

[0138] For tongue diagnosis characteristics Its contribution The calculation formula is:

[0139]

[0140] in, Indicating the characteristics of tongue diagnosis Data related to the final TCM syndrome diagnosis covariance, and These respectively represent the characteristics of tongue diagnosis. The variance of the final TCM syndrome diagnosis data D; Indicating the characteristics of tongue diagnosis Importance score in the classification of tongue diagnosis pathological patterns (can be obtained through feature selection algorithms or model interpretability methods); and It is the adjustment coefficient, and + =1, used to balance the impact of the correlation between the feature and the diagnostic result and the importance of the feature itself on the contribution;

[0141] Similarly, for facial diagnostic features Its contribution The calculation formula is:

[0142]

[0143] in Indicates facial features Data related to the final TCM syndrome diagnosis covariance, Indicates facial features The variance; Indicates facial features Importance score in the correlation analysis of pathological features during face diagnosis; and It is the adjustment coefficient, and + =1;

[0144] For pulse diagnosis characteristics Its contribution The calculation formula is:

[0145]

[0146] in, Indicating pulse diagnosis characteristics Data related to the final TCM syndrome diagnosis covariance, Indicating pulse diagnosis characteristics The variance; Indicating pulse diagnosis characteristics The importance score in the process of identifying pathological features in pulse diagnosis; and It is the adjustment coefficient, and + =1; Finally, by combining the characteristic contributions of tongue diagnosis, facial diagnosis, and pulse diagnosis, the comprehensive contribution weight of each diagnostic module to the final TCM syndrome diagnosis result is obtained. :

[0147]

[0148]

[0149]

[0150] And on Perform normalization to make it meet the requirements. + + =1.

[0151] This design allows the normalized weights to be used to more reasonably integrate the identification results of tongue diagnosis, facial diagnosis, and pulse diagnosis in the multi-source data fusion decision module, in order to generate more accurate TCM syndrome diagnosis results.

[0152] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A TCM intelligent diagnostic system based on multi-source data fusion and AI technology, characterized in that: include: The tongue image acquisition and classification unit is used to acquire patients' tongue image data through imaging equipment, extract features from the tongue image data to obtain a tongue image feature dataset; and classify the tongue image pathological patterns of the tongue image feature dataset based on a deep convolutional neural network to obtain tongue diagnosis identification results data. The facial diagnosis feature association analysis unit is used to input the tongue diagnosis identification result data into the multimodal data fusion engine, collect the patient's facial data, extract features from the facial data to obtain a facial diagnosis feature dataset; combine the tongue diagnosis identification result data to dynamically assign weights to the facial diagnosis feature dataset, and perform facial diagnosis pathological feature association analysis based on a graph attention network to obtain facial diagnosis identification result data. The pulse diagnosis feature identification and processing unit is used to input the facial diagnosis identification result data into the pulse diagnosis and consultation joint analysis module, collect the patient's pulse feature data, perform time-frequency joint analysis and waveform feature clustering on the pulse signal to obtain the pulse diagnosis feature dataset; combine the facial diagnosis identification result data with the pulse diagnosis feature dataset to perform syndrome pulse mapping matching, and perform pulse diagnosis pathological feature identification based on a bidirectional long short-term memory network to obtain the pulse diagnosis identification result data. The three-diagnosis fusion decision push unit is used to input the tongue diagnosis identification result data, face diagnosis identification result data and pulse diagnosis result data into the multi-source data fusion decision module, construct the three-diagnosis information association matrix and perform feature weight allocation; Based on graph neural networks, information fusion and reasoning of the four diagnostic methods are used to generate TCM syndrome diagnosis results data and push them to terminal devices to perform TCM intelligent diagnosis and auxiliary decision-making. Let the tongue diagnosis identification results data be The facial diagnosis and identification results data are The pulse diagnosis results data are as follows ,in , , These represent the number of feature dimensions in the identification results data for tongue diagnosis, facial diagnosis, and pulse diagnosis, respectively. Define the characteristics of tongue diagnosis The contribution to the final diagnostic result is Facial features The contribution to the final diagnostic result is Pulse diagnosis characteristics The contribution to the final diagnostic result is ,in ; The contribution of each feature is calculated based on the correlation between the feature and the final syndrome diagnosis result, as well as the importance of the feature in its respective diagnostic module, and is calculated in the following way: For tongue diagnosis characteristics Its contribution The calculation formula is: in, Indicating tongue diagnosis characteristics Data related to the final TCM syndrome diagnosis covariance, and These respectively represent the characteristics of tongue diagnosis. The variance of the final TCM syndrome diagnosis data D; Indicating tongue diagnosis characteristics The importance score in the classification of tongue diagnosis pathological patterns can be obtained through feature selection algorithms or model interpretability methods. and It is the adjustment coefficient, and + =1, used to balance the impact of the correlation between the feature and the diagnostic result and the importance of the feature itself on the contribution; Similarly, for facial diagnostic features Its contribution The calculation formula is: in Indicates facial features Data related to the final TCM syndrome diagnosis covariance, Indicates facial features The variance; Indicates facial features Importance score in the correlation analysis of pathological features during face diagnosis; and It is the adjustment coefficient, and + =1; For pulse diagnosis characteristics Its contribution The calculation formula is: in, Indicating pulse diagnosis characteristics Data related to the final TCM syndrome diagnosis covariance, Indicating pulse diagnosis characteristics The variance; Indicating pulse diagnosis characteristics The importance score in the process of identifying pathological features in pulse diagnosis; and It is the adjustment coefficient, and + =1; Finally, by combining the characteristic contributions of tongue diagnosis, facial diagnosis, and pulse diagnosis, the comprehensive contribution weight of each diagnostic module to the final TCM syndrome diagnosis result is obtained. : And on Perform normalization to make it meet the requirements. The normalized weights are used to more reasonably integrate the identification results of tongue diagnosis, face diagnosis, and pulse diagnosis in the multi-source data fusion decision module, so as to generate more accurate TCM syndrome diagnosis results.

2. The TCM intelligent diagnostic system based on multi-source data fusion and AI technology according to claim 1, characterized in that: The facial data features include micro-expression recognition and blood circulation distribution, and the patient's facial data includes facial three-dimensional point cloud data and thermal imaging data.

3. The TCM intelligent diagnostic system based on multi-source data fusion and AI technology according to claim 2, characterized in that, Tongue image data of patients is acquired through imaging equipment to obtain a tongue image feature dataset, including: The patient's tongue image data was acquired using an imaging device. The tongue image data included the reflectance spectrum information of the tongue surface and the transmission spectrum information of the subcutaneous tissue. The patient's tongue image data was then processed by a dynamic exposure compensation algorithm to equalize the brightness, resulting in the original tongue image data. The original tongue image data is input into the three-dimensional tongue modeling module, and the structured light projection technology is used to obtain the micron-level morphological data of the tongue surface. The dynamic deformation correction of the tongue is achieved by combining the elastic deformation registration algorithm. The three-dimensional geometric model of the tongue is constructed by multi-view image stitching technology to obtain the three-dimensional tongue image data. Three-dimensional tongue image data is input into a multi-dimensional feature extraction engine, and an adversarial generative network is used to enhance the tongue image data to generate a simulated tongue image dataset. Based on the improved U-Net semantic segmentation network, the tongue image is divided into regions to obtain a structured tongue image feature dataset. The structured tongue image feature dataset is input into the feature optimization and fusion module. The self-attention mechanism is used to assign weights to features of different dimensions. Cross-modal feature association analysis is achieved through graph convolutional networks. A contrastive learning strategy is used to eliminate interference from individual differences, resulting in a standardized tongue image feature vector set.

4. The TCM intelligent diagnostic system based on multi-source data fusion and AI technology according to claim 3, characterized in that, Based on a deep convolutional neural network, tongue image feature datasets are classified into pathological patterns to obtain tongue diagnosis identification results, including: The standardized tongue image feature vector set is input into the multi-scale feature perception module to extract local and global features of tongue image texture, color and edge to obtain a multi-scale feature map of the tongue image; the channel attention mechanism is used to dynamically assign weights to the multi-scale feature map of the tongue image to generate an enhanced feature map. The enhanced feature map is input into the dynamic path selection network; the contribution of each path to the classification of the current tongue image sample is evaluated in real time through a reinforcement learning agent, resulting in customized feature representation data. Customized feature representation data is input into the cross-modal association learning module, and semantic association between tongue image features and pathological syndromes is constructed through graph neural network. A contrastive learning strategy is adopted to maximize the feature similarity of positive sample pairs, while a memory bank mechanism is used to store historical sample features to enhance the model's ability to identify rare pathological patterns, resulting in semantically enhanced tongue image feature embedding data. By embedding semantically enhanced tongue image features into the data input uncertainty-aware classification head, tongue diagnosis identification results are obtained.

5. The TCM intelligent diagnostic system based on multi-source data fusion and AI technology according to claim 4, characterized in that, The tongue diagnosis results are input into a multimodal data fusion engine to obtain a facial diagnosis feature dataset, including: The tongue diagnosis results data are input into the prior knowledge injection module of the multimodal data fusion engine to obtain the tongue diagnosis prior heatmap; A spectral 3D facial scanner was used to acquire dynamic micro-expression video streams of patients’ facial structured light data and texture reflectance spectral data to generate multi-dimensional raw facial data. The tongue diagnosis prior heat map was superimposed with the facial scan data, and the network was guided to focus on highly correlated areas through a spatial attention mechanism to generate facial input data. Facial input data is input into a dynamic feature extraction network, which includes a static branch structure and a dynamic branch structure. The dynamic feature extraction network performs semantic alignment and fusion through a cross-modal interaction module to obtain joint feature representation data for tongue diagnosis and facial diagnosis. The combined feature representation data of tongue diagnosis and facial diagnosis is input into the feature decoupling and reconstruction module, which includes: Feature decoupling is achieved by separating pathological components and individual difference noise in facial features through a categorical conditional variational autoencoder. The reconstruction module dynamically adjusts the feature reconstruction intensity by combining the pathological severity score in the tongue diagnosis results, and generates a standardized facial diagnosis feature vector set.

6. The TCM intelligent diagnostic system based on multi-source data fusion and AI technology according to claim 5, characterized in that, By combining the tongue diagnosis identification data with the facial diagnosis feature dataset and dynamically assigning weights, the facial diagnosis identification data is obtained, including: The tongue diagnosis results are input into the dynamic weight generation module. A pathological association knowledge graph is constructed through the TCM tongue surface correspondence theory, and tongue diagnosis keywords are mapped to facial diagnosis feature weight parameters. At the same time, a reinforcement learning framework is introduced, with the maximization of mutual information between facial diagnosis features and tongue diagnosis results as the reward function, to obtain tongue diagnosis weight allocation strategy data. A facial imaging system is used to collect multimodal facial data of patients to generate a raw facial diagnosis dataset. The tongue diagnosis weight parameters are spatially superimposed with the facial data, and the network is guided to focus on high-weight areas through an attention mechanism to generate tongue diagnosis-enhanced facial diagnosis input data. The facial diagnosis input data enhanced by tongue diagnosis is input into a cross-modal feature extraction network to obtain dynamic facial functional features; The cross-modal feature extraction network integrates features through a cross-modal interaction module and adjusts feature weights based on the pathological degree of tongue diagnosis to obtain secondary weighted facial diagnosis feature data. The weighted facial diagnosis feature data is input into a dynamic graph attention network to obtain the patient's facial and tongue diagnosis heterogeneity graph: A multi-head attention mechanism is used to propagate pathological information on the heteromorphic map of the patient's face and tongue diagnosis, focusing on strengthening the information flow between highly correlated nodes of face and tongue diagnosis, and generating pathological correlation feature data of face diagnosis. The pathological correlation feature data is input into the decoupled and reconstructed diagnostic module to generate standardized facial diagnosis and identification results data.

7. The TCM intelligent diagnostic system based on multi-source data fusion and AI technology according to claim 6, wherein the facial diagnosis identification results data are input into the pulse diagnosis and questioning joint analysis module to obtain a pulse diagnosis feature dataset, characterized in that, include: The facial diagnosis results are input into the pulse diagnosis and consultation joint analysis module. A knowledge graph of syndrome and pulse correlation is constructed through the TCM theory of facial and pulse correspondence. Facial diagnosis keywords are mapped to pulse diagnosis feature weight parameters. A reinforcement learning framework is introduced, with the maximization of mutual information between pulse features and facial diagnosis results as the reward function, to generate initial weight allocation strategy data. A wearable pulse sensor array is used, combined with initial weight allocation strategy data, to dynamically adjust the sampling strategy and generate raw pulse signal data; The original pulse signal data is input into the time-frequency analysis module, and the pulse harmonic components are extracted using adaptive wavelet transform. The characteristic frequency bands are then highlighted by weighting the facial diagnosis weight parameters. Subsequently, the pulse signal is decomposed into multiple intrinsic mode functions using variational mode decomposition, and a pulse diagnosis feature dataset is generated by using a spectral clustering algorithm guided by facial diagnosis syndrome types.

8. The TCM intelligent diagnostic system based on multi-source data fusion and AI technology according to claim 7, which combines facial diagnosis identification results data with pulse diagnosis feature dataset to perform syndrome pulse image mapping and matching to obtain pulse diagnosis identification results data, is characterized in that... include: The facial diagnosis results are input into a dynamic weight allocation engine to extract the correlation weight matrix between facial meridian energy distribution features and tongue diagnosis syndrome labels. Based on this weight matrix, the pulse diagnosis feature dataset is recalibrated to generate a pulse energy distribution map. Based on the syndrome types in the facial diagnosis results data, a search and matching is performed in the pre-built syndrome pulse association knowledge base to obtain typical pulse feature patterns. The typical pulse feature patterns are then compared with the initial pulse diagnosis feature dataset to obtain a subset of pulse diagnosis features. A subset of pulse diagnosis features is input into a bidirectional long short-term memory network model trained with a large amount of pulse diagnosis data. Through the forward and backward propagation process of the bidirectional long short-term memory network, deep learning and feature mining are performed on the subset of pulse diagnosis features to obtain the pulse diagnosis feature vector. The similarity calculation and matching analysis are performed between the pulse diagnosis feature vector and the pre-set standard template of pulse feature corresponding to different syndromes. The matching results are then comprehensively evaluated and adjusted by combining the syndrome weight information in the face diagnosis identification results data to obtain the pulse diagnosis identification results data.

9. The TCM intelligent diagnostic system based on multi-source data fusion and AI technology according to claim 8, wherein tongue diagnosis identification results data, facial diagnosis identification results data, and pulse diagnosis identification results data are input into the multi-source data fusion decision module to generate TCM syndrome diagnosis results data, characterized in that, include: The results of tongue diagnosis, face diagnosis, and pulse diagnosis are input into the multi-source data fusion decision module. The pathological keywords of each diagnosis result are analyzed through the TCM three-diagnosis association knowledge graph to generate a structured feature vector set. Based on the structured feature vector set, a multi-dimensional correlation analysis algorithm is used to generate the correlation matrix of the three diagnostic methods information: The three-diagnosis information association matrix includes horizontal association, vertical reinforcement and conflict resolution. The three-diagnosis information association matrix is ​​adjusted by posterior probability through a Bayesian network to dynamically optimize the weight allocation and obtain a three-dimensional weight matrix. The three-dimensional weight matrix and the structured feature vector set are input into the graph neural network fusion engine to construct the patient's four diagnostic information graph: the four diagnostic information graph includes node construction, edge dynamic evolution and integration, and attention propagation; Traditional Chinese medicine syndrome diagnosis data is generated through a graph classification layer. The TCM syndrome diagnosis data includes syndrome type, confidence level, and cross-modal evidence chain.

Images