Body-disease-syndrome multi-dimensional dynamic identification and early warning system and method thereof

Abstract

This invention relates to the field of reinforcement learning technology, specifically to a multidimensional dynamic identification and early warning system and method for body-disease-syndrome. In this invention, the Qi and Blood tensor is extracted by scheduling the coordinates of the particle set associated with belief state through a long short-term memory network, achieving a high-optimal path selection gain for temporal pathological evolution and improving the stability of the scalar within the action evaluation numerical matrix. By subtracting predicted values ​​through a convolutional neural network to construct a phlegm-dampness error matrix and calibrating the probability floating-point number, the invention achieves precise capture of sparse high-risk pathological samples and iterative step-size directional optimization gains, breaking through the bottleneck of conventional experience playback and mean square error easily falling into local suboptimal solutions. By comparing syndrome differentiation and simulated trajectories to obtain classification constants and merging Qi and Blood to construct the target tensor, the invention achieves rapid convergence gains of the multi-level diagnosis and treatment strategy network weight matrix, avoiding the risk of gradient explosion in backpropagation, thereby improving the accuracy of intervention to block the evolution of unbalanced constitution into severe illness and significantly reducing the graphics processing unit memory overhead.

Description

Technical Field

[0001] This invention relates to the field of reinforcement learning technology, and in particular to a multidimensional dynamic identification and early warning system and method for body-disease-symptom. Background Technology

[0002] Reinforcement learning comprises two specific computational architectures: a proximal policy optimization algorithm and a deep deterministic policy gradient algorithm. This avoids encompassing too many generic models. The underlying operations include constructing a closed loop of a TCM constitution state space tensor, an intervention action space vector, and a TCM clinical efficacy reward feedback matrix. The discount factor hyperparameter constant is set to 0.99. The scheduling tensor processing unit performs temporal difference updates using the Bellman equation. Multiple batch sample tensors containing 256 dimensions are extracted using an empirical replay buffer. An adaptive moment estimation optimization operator is instructed to reduce the mean square error loss of the multi-layer neural network until it reaches a constant limit of 0.005. Finally, the network weight matrix of the TCM diagnosis and treatment policy function, which meets the set convergence conditions, is output.

[0003] The TCM constitution-disease-syndrome multidimensional dynamic identification and early warning system is specifically based on a partially observable Markov decision process and integrates a computational hardware cluster of TCM classical syndrome differentiation and treatment ontology reasoning computation graph. The system construction objectives include mapping a three-dimensional transition probability distribution matrix of biased constitution transformation nodes, TCM disease classification nodes, and syndrome evolution nodes in the graphics processing unit's memory space; calculating multiple implicit TCM pathological states to map emission probability constant matrices to the observation space; and achieving multiple disease prevention and intervention effects by calculating the cumulative return floating-point number of the three-dimensional state path transition discount of constitution, disease, and syndrome; comparing the output value of the global state action value tensor with the 0.85% critical benchmark constant for constitution bias; and immediately generating multi-level classical prescriptions and acupuncture point blocking execution script code when the calculated value is higher than 15% of the benchmark, intercepting the directed graph edge weight update of high-risk phlegm-dampness constitution evolving into severe turbid toxin accumulation, and suppressing the probability constant of qi and blood imbalance transfer to within the lower limit of the 0.12% safety threshold.

[0004] The existing technical architecture heavily relies on the full mapping of the graphics processing unit's memory space to a biased three-dimensional transition probability distribution matrix for constitution transformation nodes, TCM disease classification nodes, and syndrome evolution nodes. When faced with observable Markov decision processes in high dimensions, it forces dense matrix operations, inevitably leading to hardware memory addressing overflow risks. Furthermore, using a fixed 256-dimensional numerical batch sample tensor for experience replay extraction, and assigning equal sampling weights to multiple historical trajectories, results in sparse and high-risk acute pathological state samples being submerged within a massive array of conventional physiological fluctuation data. Finally, the superimposed adaptive moment estimation optimization operator aims solely to reduce the mean square error loss of the multi-layer neural network to zero. The five constant limits can easily lead to the policy network falling into an overfitting state of high-frequency stable evolution trajectory and losing the gradient response of abnormal pathological evolution edge nodes. When the input captures the tensor of the mutation transfer from Yin deficiency state to severe internal turbidity and toxin accumulation, the conventional mean square error loss calculation and homogenized experience playback extraction mechanism will infinitely dilute multiple low-frequency and high-risk temporal differential update signals, which will lead to the output value of the global state action value tensor being delayed and lower than the critical reference constant of 0.85 for the disease of the constitution bias. This directly misses the opportunity to generate multi-level classical prescriptions and acupuncture point blocking script code, and allows the probability constant of Qi and blood imbalance to exceed the lower limit of the safety threshold of 0.12, resulting in irreversible clinical variation consequences. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology and to propose a multidimensional dynamic identification and early warning system and method for body-disease-symptom.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: the multidimensional dynamic identification and early warning system for body-disease-symptom includes: Latent pathogenesis optimization module: By using clinical four diagnostic parameters and physical constitution assessment scale scores, the physical constitution pathology array is defined, the Yin deficiency tensor and phlegm-dampness internal obstruction matrix are entered, the inner product summation value is calculated and defined as the top source level, and the belief state particle set is obtained. Dynamic Evolutionary Game Theory Module: Based on the coordinates of the particle set of the belief state, the constant is calculated by dividing the Qi and Blood tensor by the frequency through the Long Short-Term Memory Network, the boundary tensor is extracted and the constitution flow frequency is back-stacked, the first and second sequence is selected, and the action code of the first Jingfang is output. Competition flow splitting module: Based on the action encoding of the first prescription, extract the constitution tensor and split it into the spleen and stomach transportation and transformation branch and the liver qi stagnation branch. Calculate the output tensors of the two branches, add the tensor values ​​and perform scaling operation to generate the action evaluation numerical matrix. Priority playback optimization module: Based on the action evaluation numerical matrix, the predicted values ​​are subtracted by the convolutional neural network to form the phlegm and dampness error matrix, the probability floating-point numbers are calibrated and samples above the limit constant are extracted, the coefficients are multiplied by the gradient constant, and the iterative update parameter tensor is established. The adversarial feedback derivation module: Based on the action evaluation numerical matrix and the iterative update parameter tensor, it compares the dialectical and simulated trajectories to obtain the classification constant, merges the qi and blood to construct the target tensor, and alternately executes the derivative rise and fall until it drops to the boundary and stops, thus obtaining the instantaneous dynamic reward constant.

[0007] As a further aspect of the present invention, the belief state particle set includes a Yin deficiency tensor, a phlegm-dampness internal resistance matrix, and a source-level position scalar; the primary classical formula action encoding includes a boundary tensor, a constitution flow frequency value, and a core classical formula blocking intervention constant; the action evaluation numerical matrix includes a spleen and stomach transportation and transformation branch output tensor, a liver qi stagnation branch output tensor, and a tensor overflow scalar; the iterative update parameter tensor includes a probability floating-point number, a sampling multiplier coefficient, an error gradient constant, and a penalty scalar; and the instantaneous dynamic reward constant includes a classification constant, a syndrome differentiation offset scalar array, and a qi and blood information entropy parameter.

[0008] As a further aspect of the present invention, the latent pathogenesis optimization module includes: The array structure construction submodule: Based on clinical characterization parameters and body mass scale scores, it constructs a three-dimensional coordinate system of body constitution, inputs the values ​​of the Yin deficiency and dryness-heat bias tensor and extracts the intersection points of the phlegm-dampness internal obstruction matrix, aligns multiple pathological element scalars, summarizes the multidimensional pathological data architecture, and generates a multidimensional bias feature matrix. Source-level partitioning submodule: Based on the multidimensional biased feature matrix, extract multiple row and column space coordinate elements, calculate the tensor element inner product summation constant, summarize the summed values ​​to define the initial boundary points, set the source-level position scalar, match the bottom architecture exploration starting point, and obtain the belief state particle set.

[0009] As a further aspect of the present invention, the dynamic evolutionary game module includes: Qi and Blood Tensor Sampling Submodule: Based on the coordinates of the belief state particle set, deploy a long short-term memory network to cyclically extract the evolution characteristics of Qi and Blood time series, divide the space of Qi and Blood circulation tensor array, read the frequency scalar and floating-point constant of Qi deficiency Yang micro, issue an instruction to divide the tensor elements by the frequency scalar, merge the floating-point constant to reconstruct the exploration data matrix, and obtain the evolution state exploration constant. Branch frequency accumulation submodule: Based on the evolution state exploration constant, compare the internal Qi and blood boundary values ​​of the constant, extract the phlegm-dampness stagnation tensor and trace the branch path to the corresponding level coordinates, extract the constitution flow frequency value, issue an addition command to accumulate the level-related frequency value, and obtain the state backtracking value array. First-position sequence selection submodule: Based on the state backtracking numerical array, read the spatial coordinates of the elements inside the array, compare and filter the numerical values ​​of the array elements, arrange the descending position sequence of the array elements, extract the action sequence associated with the first-position value and extract the core classical prescription blocking intervention constant, and obtain the action code of the first classical prescription.

[0010] As a further aspect of the present invention, the Long Short-Term Memory (LSTM) network extracts the Qi and Blood temporal tensor sequence, initializes the hidden state tensor and cell state constant, extracts the current Qi and Blood tensor within the Qi and Blood temporal tensor sequence, executes the inner product summation instruction of the hidden state tensor and the current Qi and Blood tensor to generate an intermediate state matrix, calls the activation operator to calculate the intermediate state matrix to obtain the retention probability matrix, simultaneously combines the current Qi and Blood tensor to execute the inner product instruction to generate candidate incremental tensors, multiplies the retention probability matrix and the cell state constant to obtain the forgotten state tensor, superimposes the forgotten state tensor and the candidate incremental tensor and overwrites the cell state constant, issues a tensor mapping instruction to the overwritten cell state constant, and outputs the Qi and Blood temporal evolution feature array.

[0011] As a further aspect of the present invention, the competitive flow splitting module includes: Tensor Cutting and Splitting Submodule: Based on the action encoding of the first prescription, map the feature parameters of the constitution tensor dimension, split the spleen and stomach transportation and transformation and liver qi stagnation branches, extract the tensor related to spleen and stomach and liver qi stagnation, multiply the tensor related to spleen and stomach and liver qi stagnation and then merge the tensor product constant to obtain the split branch output array. Tensor numerical scaling submodule: Based on the split branch output array, extract independent tensor data within the array, add the values ​​of the corresponding elements of the independent tensors, introduce a proportional constant to scale the tensor sum value, define the parameter mapping boundary to truncate the tensor overflow scalar, and generate an action evaluation numerical matrix.

[0012] As a further aspect of the present invention, the priority playback optimization module includes: Bias Feature Construction Submodule: Based on the action evaluation numerical matrix, deploy a convolutional neural network pathological feature parsing architecture to read historical sample data of Qi and Blood constitution, retrieve the corresponding coordinate nodes inside the matrix, align the expected value with the actual predicted scalar, issue a subtraction instruction to calculate the absolute difference, collect the difference to construct a floating-point distribution space, and generate a phlegm-dampness transformation error array. The over-threshold screening submodule: Based on the phlegm-dampness transformation error array, it divides the hierarchical error echelon boundary distribution, reads the array elements to define the hierarchical parameters and marks the corresponding hierarchical probability floating-point values, compares the probability floating-point numbers with the boundary constant, extracts the historical tensors that are higher than the boundary constant, and obtains the over-limit constitution intervention samples. The product derivative merging submodule: Based on the over-boundary physical intervention samples, extract the sampling multiplier coefficients inside the samples and locate the coordinates of the bottom sample to read the coefficient values. Multiply the sampling coefficients with the error gradient constant, calculate the product value to form a penalty scalar, merge the penalty scalar with the pre-stored co-position parameters, and establish an iteratively updated parameter tensor.

[0013] As a further aspect of the present invention, the convolutional neural network reads historical sample data of Qi and Blood constitution, constructs a spatial array of Qi and Blood constitution, sets local convolution kernel parameters, controls the local convolution kernel to perform sliding traversal in the spatial array of Qi and Blood constitution, calculates the product constant of the same elements, summarizes the product constant to generate a hidden pathological tensor, calls a nonlinear activation operator to map the hidden pathological tensor, delineates the dimensionality reduction interval and extracts the peak constant of the interval, reduces the size of the hidden pathological tensor, and outputs a Qi and Blood pathological feature array.

[0014] As a further aspect of the present invention, the adversarial feedback derivation module includes: Trajectory comparison and classification submodule: Based on the action evaluation numerical matrix and the iterative update parameter tensor, extract the pre-stored expert dialectical trajectory tensor, retrieve the corresponding simulated evolution trajectory tensor and align the corresponding spatial coordinate elements, calculate the absolute numerical difference of the real scalar, and obtain the dialectical offset scalar array. The target tensor differentiation submodule extracts the entropy parameters of the Qi and Blood information within the array based on the dialectical offset scalar array, merges the entropy parameters with the scalar array to form a joint target feedback tensor, alternately issues element derivative increase and decrease commands, stops the operation when the boundary scalar is reached, and obtains the instantaneous dynamic reward constant.

[0015] A multidimensional dynamic identification and early warning method for body-disease-symptom, which is based on the aforementioned multidimensional dynamic identification and early warning system for body-disease-symptom, includes the following steps: S1: Based on the clinical four diagnostic parameters and the scores of the constitution assessment scale, the constitution spatial architecture is mapped and the Yin deficiency tensor value is injected. The intersection of the phlegm-dampness internal resistance matrix is ​​extracted, the tensor inner product summation constant is calculated, the top-level source level coordinate starting point is defined, and the belief state particle set is obtained. S2: Based on the coordinates of the particle set of the belief state, extract the historical time sequence features and calculate the Qi and Blood tensor divided by the frequency constant decay scalar to divide the boundary tensor. Backtrack the branch to accumulate the constitution flow frequency, sort the descending position to filter the first sequence, and output the first Jingfang action code. S3: Based on the action encoding of the first prescription, map the deep architecture constitution features and separate the tensors of spleen and stomach transportation and liver qi stagnation, multiply and add the correlation tensors and merge the product values, introduce a scaling coefficient to scale and truncate the tensors and sum the overflow scalars to generate the action evaluation numerical matrix. S4: Based on the action evaluation numerical matrix, traverse the Qi and Blood sample spatial array, align the expected scalar subtraction difference to form the phlegm and dampness error matrix, calibrate the hierarchical probability to extract the out-of-limit constant sample, multiply by the sampling coefficient and the error gradient constant, and establish the iterative update parameter tensor. S5: Based on the action evaluation numerical matrix and the iterative update parameter tensor, align the dialectical and simulated evolution trajectories, calculate the numerical difference to obtain the classification constant, merge the blood and qi parameters to splice the target tensor, alternately increase and decrease the derivative until the boundary scalar stops, and obtain the instantaneous dynamic reward constant.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, the Qi and Blood tensor is extracted by scheduling the Long Short-Term Memory Network with the coordinates of the particle set associated with the belief state. The frequency calculation constant division operation is performed and the boundary tensor is truncated and superimposed with the constitution flow frequency to achieve the gain of high-optimal path selection for temporal pathological evolution and improve the stability of the scalar within the action evaluation numerical matrix. In this invention, the phlegm-dampness error matrix is ​​constructed by subtracting predicted values ​​through a convolutional neural network, and the floating-point probability is extracted to extract the multiplication coefficients of samples above the limit constant and establish an iterative update parameter tensor with the gradient constant. This achieves accurate capture of sparse high-risk pathological samples and iterative step size optimization gain, breaking through the bottleneck of conventional experience playback and mean square error that easily falls into local suboptimal solutions. In this invention, by comparing and distinguishing and simulating trajectories, the classification constant is obtained and the target tensor is constructed by merging Qi and Blood. The derivative rises and falls alternately until the boundary is reached and the instantaneous dynamic reward constant is obtained. This achieves the rapid convergence gain of the weight matrix of the multi-level diagnosis and treatment strategy network, avoids the risk of gradient explosion in backpropagation, and thus improves the accuracy of intervention to prevent the evolution of skewed constitution into severe disease and significantly reduces the memory overhead of graphics processing unit. Attached Figure Description

[0017] Figure 1 This is a system flowchart of the present invention; Figure 2 This is a schematic diagram of the method steps of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] Example 1 Please see Figure 1 The present invention provides a technical solution: a multi-dimensional dynamic identification and early warning system for body-disease-symptom.

[0020] Latent pathogenesis optimization module: By using clinical four diagnostic parameters and physical constitution assessment scale scores, the physical constitution pathology array is defined, the Yin deficiency tensor and phlegm-dampness internal obstruction matrix are entered, the inner product summation value is calculated and defined as the top source level, and the belief state particle set is obtained. Dynamic Evolutionary Game Theory Module: Based on the coordinates of the belief state particle set, the constant is calculated by dividing the Qi and Blood tensor by the frequency through the Long Short-Term Memory Network, the boundary tensor is truncated and the constitution flow frequency is back-stacked, the first and second sequence is selected, and the action code of the first Jingfang is output. Competition flow splitting module: Based on the action encoding of the first prescription, extract the constitution tensor and split it into the spleen and stomach transportation and transformation branch and the liver qi stagnation branch. Calculate the output tensors of the two branches, add the tensor values ​​and perform scaling operation to generate the action evaluation numerical matrix. Priority playback optimization module: Based on the action evaluation numerical matrix, the predicted values ​​are subtracted by the convolutional neural network to form the phlegm and dampness error matrix. The probability floating-point numbers are calibrated and samples above the limit constant are extracted. The coefficients are multiplied with the gradient constant to establish an iterative update parameter tensor. The adversarial feedback derivation module: Based on the action evaluation numerical matrix and the iteratively updated parameter tensor, the classification constant is obtained by comparing the dialectical and simulated trajectories, the target tensor is constructed by merging the qi and blood, and the derivative rise and fall are alternately executed until it stops at the boundary, thus obtaining the instantaneous dynamic reward constant.

[0021] The belief state particle set includes the Yin deficiency tensor, the phlegm-dampness internal resistance matrix, and the source-level position scalar. The primary classical formula action encoding includes the boundary tensor, the constitution flow frequency value, and the core classical formula blocking intervention constant. The action evaluation numerical matrix includes the output tensor of the spleen and stomach transportation and transformation branch, the output tensor of the liver stagnation and qi stagnation branch, and the tensor overflow scalar. The iterative update parameter tensor includes the probability floating-point number, the sampling multiplier coefficient, the error gradient constant, and the penalty scalar. The instantaneous dynamic reward constant includes the classification constant, the syndrome differentiation offset scalar array, and the qi and blood information entropy parameter.

[0022] The latent pathogenesis optimization module includes: The array structure construction submodule: Based on clinical characterization parameters and body mass scale scores, it constructs a three-dimensional coordinate system of body constitution, inputs the values ​​of the Yin deficiency and dryness-heat bias tensor and extracts the intersection points of the phlegm-dampness internal obstruction matrix, aligns multiple pathological element scalars, summarizes the multidimensional pathological data architecture, and generates a multidimensional bias feature matrix. Source-level partitioning submodule: Based on the multi-dimensional biased feature matrix, extract multiple row and column space coordinate elements, calculate the tensor element inner product summation constant, summarize the summed values ​​to define the initial boundary point, set the source-level position scalar, match the exploration starting point of the underlying architecture, and obtain the belief state particle set; The array structure construction submodule establishes a mapping scheme for Qi and Blood tensors based on clinical representation parameters and body mass index scores, associating the original four diagnostic methods and scale data. It retrieves a table of correspondence between TCM concepts and algorithm variables, extracts the hidden state tensor dimension constant corresponding to the Qi and Blood imbalance concept and the feature boundary scalar corresponding to the Spleen and Stomach transportation concept, and retrieves a data field mapping table. It maps the pulse frequency field within the clinical representation parameters to a single-precision floating-point tensor and the tongue color pixel matrix field to an unsigned integer tensor. It extracts the pulse frequency constant and tongue color pixel matrix within the clinical representation parameters, merges the symptom rating scalar within the body mass index scores, and inputs them into a multilayer perceptron network to perform one-hot encoding and nonlinear mapping operations, generating multiple three-dimensional Qi and Blood tensors. It uses a Gaussian mixture model algorithm and a tensor singular value decomposition algorithm to construct the three-dimensional coordinates of the body mass index space. It calls the tensor decomposition operator, inputs the array of multiple representation parameters, and performs matrix operations. It sets a singular value truncation parameter constant of 3 to extract the left and right singular tensors, and truncates the first three largest terms. Singular value scalars are used as basis vectors. The number of components in the Gaussian mixture distribution is set to 5. The covariance matrix type parameter is set to a full matrix constant. The preset empirical weight constant is read as 0.8. The above preset operation is to input 10,000 historical medical records, call the linear regression operator to calculate floating-point numbers and write them to the fixed address segment of the register. The upper limit of the expected maximum iteration step is set to 500 times. The tolerance threshold floating-point number parameter is input as 0.0001. The instruction program executes iterative optimization in a loop until the termination condition is triggered. The coordinates of the center position of the output spatial distribution are calculated. The values ​​of the Yin deficiency and dryness-heat bias tensor are entered to extract the intersection of the phlegm-dampness internal resistance matrix. The floating-point number vector of the intersection position of the Yin deficiency tensor and the phlegm-dampness matrix inside the above spatial coordinates is extracted. The tensor dot product operator is called to multiply by the spatial scaling constant 1.5. The scalars of multiple pathogenesis elements are aligned. The feature splicing interface offset constant is set to 0. The pooling downsampling operator is called to set the sliding window scale to 2 and the step size constant to 2. Dimensionality reduction processing is performed to summarize the multidimensional pathological data architecture and generate a multidimensional bias feature matrix. The source-level sub-module is divided as follows: Based on a multi-dimensional biased feature matrix, density clustering and hidden Markov model algorithms are used to extract multiple row and column space coordinate elements. The clustering density detection operator is called to read the multi-dimensional feature array. A floating-point constant parameter of 0.5 is configured for the neighborhood radius, and an integer parameter of 15 is configured for the minimum number of objects in the neighborhood of the core node. The instruction calculates the discrete coordinate vectors of the core point's interconnectedness, constructs a connected graph to extract boundary contour node coordinates, imports a hidden state quantity parameter of 8, imports an observed state quantity parameter of 12, and reads the initial array with a specified emission probability. Specifically, the above specified operations involve the main thread sending a memory access instruction to the coprocessor and passing in a hexadecimal starting address code constant to retrieve the... In response to data block operations, the load transfer probability distribution array is run before and after the Viterbi decoding operator outputs probability floating-point numbers, and the summation of tensor element inner product constants is calculated to define the initial boundary points. The tensor dot product operator is called to traverse and extract elements and associated weight vectors, and multiply-add and accumulate instructions are executed. The multiple summation constants are compared with the judgment threshold parameter 1.2 to obtain the Boolean classification scalar, and data dimension constants below the boundary threshold are removed. The source level position scalar is set, and the coordinates of high-dimensional feature points are mapped to the two-dimensional decision grid plane. The level label identifier constant is assigned 1, the underlying architecture exploration starting point is matched, and the first address pointer constant is issued to the central processing unit thread to start the depth traversal search and obtain the belief state particle set.

[0023] The dynamic evolutionary game module includes multiple reinforcement learning optimization branches, and sets up a reinforcement learning framework to undertake the core computational tasks, specifically constructing a Markov decision process, calculating the state-action value tensor, and maximizing the cumulative reward floating-point number for blocking the disease-causing evolutionary path. The module includes: Qi and Blood Tensor Sampling Submodule: Based on the coordinates of the belief state particle set, deploy a long short-term memory network to cyclically extract the evolutionary features of Qi and Blood, divide the space of Qi and Blood circulation tensor array, read the frequency scalar and floating-point constant of Qi deficiency and Yang micro, issue instructions to divide the tensor elements by the frequency scalar, merge the floating-point constants to reconstruct the exploration data matrix, and obtain the evolutionary state exploration constant. Branch Frequency Accumulation Submodule: Based on the evolution state exploration constant, compare the internal Qi and Blood boundary values ​​of the constant, extract the phlegm-dampness stagnation tensor and trace the branch path to the corresponding level coordinates, extract the constitution flow frequency value, issue an addition command to accumulate the level-related frequency value, and obtain the state backtracking numerical array. First-position sequence selection submodule: Based on the state backtracking numerical array, read the spatial coordinates of the elements inside the array, compare and filter the numerical values ​​of the array elements, arrange the descending position sequence of the array elements, extract the action sequence associated with the first-position value and extract the core classical formula blocking intervention constant, and obtain the action code of the first classical formula. The Qi and Blood Tensor Sampling Submodule: Based on the coordinates of the belief state particle set, it uses a Long Short-Term Memory (LSTM) network algorithm and tensor broadcast division instructions to cyclically extract the temporal evolution features of Qi and Blood. The LSM network is set to undertake the core computational task of capturing the nonlinear dynamic flow constant of the Qi and Blood tensor across time steps. The scheduling network initialization operation configures the hidden layer node dimension parameter to 128, sets the sequence time step parameter to 24, and assigns initial values ​​to the input gate, forget gate, output gate, weight bias constant matrix. Specifically, this initialization operation involves calling a Gaussian distribution pseudo-random generator to generate a floating-point array with a mean constant of 0 and a variance constant of 0.1, covering the memory block address. The input tensor sequence... The cell state update instruction is triggered to output the hidden state vector of the current time step in a loop. Multiple state vectors are summarized to divide the Qi and Blood circulation tensor array space. The tensor slicing operator is called to set the starting index constant 0 and the truncation length constant 64 to execute the memory address continuous read instruction. The internal cache of the register is retrieved, and the frequency scalar and floating-point constant of Qi deficiency and Yang micro are retrieved. The floating-point constant value is set to 0.05. The instruction is issued to activate the hardware floating-point arithmetic unit to traverse the tensor elements and divide them by the frequency scalar. The tensor splicing operator is called to set the splicing dimension parameter 1 to merge and calculate the output quotient value and floating-point constant. The original memory pointer address is overwritten to reconstruct the exploration data matrix and obtain the evolution state exploration constant. Branch frequency accumulation submodule: Based on the evolutionary state exploration constant, it uses the Monte Carlo tree search algorithm and tensor mask comparison operator to compare the constant internal Qi and blood boundary values, and imports the preset Qi and blood boundary safety threshold parameter of 0.85. The above preset operation specifically involves reading the floating-point fields inside the disk's underlying configuration file during the system startup phase and loading them into read-only memory to perform address binding operations. It then calls the tensor mask comparison operator to compare the scalar inside the evolutionary state exploration constant with the boundary threshold parameter bit by bit, outputting a Boolean mask matrix. Finally, it executes the mask dot product instruction to filter outbound elements and extract phlegm and dampness. The stagnation tensor is used to run the tree search node expansion function with an exploration confidence hyperparameter constant of 1.414. The simulation depth limit is set to an integer upper limit of 50. The above tensor constant is injected into the reverse topology computation graph to trace the branch path to the corresponding level coordinates. The frequency values ​​of the associated physical flow of the level node are extracted. A multi-threaded addition instruction is issued to the arithmetic logic unit to extract the current level associated frequency value and the child node feedback frequency value from the register cache and perform a 64-bit high-precision floating-point accumulation operation. The summation of the traversal branch node summation constant is written to the heap memory space to obtain the state backtracking numerical array. The first-order sequence selection submodule, based on a state-backtracking numerical array, employs a dual-axis fast sorting algorithm and a hash mapping addressing algorithm to read the spatial coordinates of elements within the array. It calls a memory-direct read operator to lock the array's first address pointer, scans and extracts the spatial coordinates of all elements, activates the sorting processor core to compare and filter the numerical values ​​of array elements, sets a pivot selection logic pointer to point to the memory address of the current array interval, configures a recursive segmentation layer limit parameter of 100, alternately moves the left and right endpoints, compares the floating-point numbers at corresponding positions, performs element position swapping and memory overwriting operations, sorts the array elements into a descending order sequence, extracts the first-order index constant of 0 to truncate the first-order value associated with the action sequence, initializes the hash table memory allocation constant of 1024, and loads pre-specified intervention constant key-value pairs. Specifically, during the compilation phase, the key-value data is hard-coded and embedded into the executable program by associating with external medical configuration constant header files. The data segment operation constructs a scheme for generating action codes from classical prescriptions and intervention plans. It consults the classical prescription and intervention plan coding rule table stored in read-only memory, converts the drug compatibility dosage constants within the classical prescriptions into eight-bit binary action masks according to the drug mapping rules in the first column of the rule table, and converts the three-dimensional coordinate floating-point numbers of acupoints within the acupuncture intervention plan into a hexadecimal fixed-point array according to the acupoint mapping rules in the second column of the rule table. It extracts the drug compatibility dosage constants within the classical prescriptions and the three-dimensional coordinate floating-point numbers of acupoints within the acupuncture intervention plan, inputs them into a hash mapping function to calculate the corresponding hexadecimal values ​​of the strings, maps the corresponding hexadecimal values ​​to a continuous action space coordinate array to generate an action code tensor, calls the hash mapping function to calculate the hash value of the action sequence string, compares the hash bucket pointers to extract the corresponding core classical prescription blocking intervention constants, packages the first and second sequence identifiers and intervention constants into the output buffer, and obtains the action code of the first classical prescription.

[0024] The Long Short-Term Memory (LSTM) network extracts the Qi and Blood temporal tensor sequence, initializes the hidden state tensor and cell state constant, extracts the current Qi and Blood tensor within the Qi and Blood temporal tensor sequence, executes the inner product summation instruction of the hidden state tensor and the current Qi and Blood tensor to generate an intermediate state matrix, calls the activation operator to calculate the intermediate state matrix to obtain the retention probability matrix, simultaneously combines the current Qi and Blood tensor to execute the inner product instruction to generate candidate incremental tensors, multiplies the retention probability matrix and the cell state constant to obtain the forgotten state tensor, superimposes the forgotten state tensor and the candidate incremental tensor and overwrites the cell state constant, issues tensor mapping instructions for the overwritten cell state constant, and outputs the Qi and Blood temporal evolution feature array.

[0025] The competitive flow splitting module includes: Tensor Cutting and Splitting Submodule: Based on the action encoding of the first prescription, the module maps the feature parameters of the constitution tensor dimension, splits the spleen and stomach transportation and transformation and liver qi stagnation branches, extracts the tensor related to spleen and stomach and liver qi stagnation, multiplies the tensor related to spleen and stomach and liver qi stagnation and then merges the tensor product constant to obtain the splitting branch output array. Tensor numerical scaling submodule: Based on the split branch output array, extract the independent tensor data inside the array, add the values ​​of the corresponding elements of the independent tensors, introduce a proportional constant to scale the tensor sum value, define the parameter mapping boundary to truncate the tensor overflow scalar, and generate the action evaluation numerical matrix. Tensor Slicing Submodule: Based on the action encoding of the first-line formula, it uses a multilayer perceptron feature mapping algorithm and a tensor slice stripping operator to map the body tensor dimension feature parameters. It calls the fully connected layer network initialization instructions to configure the input layer neuron node count as a constant of 256, configures the hidden layer dimension expansion parameter as 512, sets the nonlinear activation function type parameter to a linear rectifier operator, and loads a pre-specified weight bias matrix. Specifically, during model initialization, it reads the local disk training solidified weight binary file and loads it into a pre-specified fixed address block in the graphics processing unit's memory. It injects the encoded data block into the forward propagation calculation channel, executes the matrix dot product instruction, and outputs a high-dimensional feature array. The tensor slice operator is instructed to set the axial dimension parameter of 1, and configures... The left-side starting offset pointer constant is 0 and the truncation step constant is 128. The spleen and stomach function and liver qi stagnation branches are divided. An independent continuous data block addressing instruction is issued to the hardware memory controller. The first addressing bus is loaded to extract the floating-point number of spleen and stomach characteristics stored in the front register. The second addressing bus is loaded to extract the floating-point number of liver qi stagnation status stored in the back register. The tensor associated with spleen and stomach and liver qi stagnation is extracted. The Hadamard product multiplication operator is used to perform point-to-point operation of the tensor. The floating-point multiplier core inside the arithmetic logic unit is activated. The memory address values ​​of the two tensors are read and double-precision floating-point multiplication instruction is executed. After multiplying the tensor associated with spleen and stomach and liver qi stagnation, the tensor splicing function is called to set the merging axis parameter to 0. The merging tensor product constant is overwritten to the continuous storage space of the cache data stack to obtain the split branch output array. The tensor value scaling submodule, based on the split-branch output array, employs a multi-dimensional tensor traversal addressing algorithm and tensor same-position addition instructions to extract independent tensor data within the array. It configures an outer loop boundary constant parameter of 256 to control the batch dimension traversal steps and an inner loop boundary constant parameter of 128 to control the feature dimension addressing steps. It locks the array's starting address data pointer, drives the pointer to slide byte-by-byte to extract independent tensor data within the array, calls the accumulator register inside the central processing unit, executes floating-point addition instructions to add the values ​​of the same-position elements of the independent tensors, temporarily stores the summation constant in the L1 cache high-speed data area, and uses a weighted scaling factor control algorithm to introduce a preset scaling constant to scale the tensor summation value. Specifically, during system startup, the main process reads empirical floating-point values ​​from the configuration file and forces them to be converted to constants before writing them. The global variable data dictionary memory segment is operated on. The preset proportional constant value of 0.01 is extracted. A multiplication instruction is issued to multiply the cached accumulation summation constant by the scaling constant 0.01 one by one to output the scaled result constant. The amplitude truncation operator is used to define the parameter mapping boundary and truncate the tensor overflow scalar. The upper threshold constant of 10.5 and the lower threshold constant of -10.5 are imported. The scaled result constant is compared with the upper and lower thresholds. If the current result constant exceeds the upper threshold constant of 10.5, it is directly overwritten with the upper threshold constant of 10.5. If the current result constant is lower than the lower threshold constant of -10.5, it is forcibly replaced with the lower threshold constant of -10.5. The overflow difference data block is discarded and the tensor overflow scalar is truncated. All legal floating-point constants are encapsulated and written into the two-dimensional static storage architecture space to generate the action evaluation numerical matrix.

[0026] The priority playback optimization module includes: Bias Feature Construction Submodule: Based on the action evaluation numerical matrix, deploy a convolutional neural network pathological feature parsing architecture to read historical sample data of Qi and Blood constitution, retrieve the corresponding coordinate nodes inside the matrix, align the expected value with the actual predicted scalar, issue a subtraction instruction to calculate the absolute difference, collect the difference to construct a floating-point distribution space, and generate a phlegm-dampness transformation error array. The over-threshold screening submodule: Based on the phlegm-dampness transformation error array, it divides the hierarchical error echelon boundary distribution, reads the array elements to define the hierarchical parameters and marks the corresponding hierarchical probability floating-point values, compares the probability floating-point numbers with the boundary constant, extracts the historical tensor above the boundary constant, and obtains the over-boundary constitution intervention samples. Product differentiation and merging submodule: Based on the over-boundary physical intervention samples, extract the sampling multiplier coefficients inside the samples and locate the coordinates of the bottom sample to read the coefficient values. Multiply the sampling coefficients with the error gradient constant, calculate the product value to form a penalty scalar, merge the penalty scalar with the pre-stored co-position parameters, and establish an iteratively updated parameter tensor. The biased feature construction submodule, based on the action evaluation numerical matrix, uses a convolutional neural network feature extraction algorithm and an absolute difference tensor subtraction operator to read historical sample data of Qi and blood constitution. The convolutional neural network is configured to perform local receptive field convolution operations and reduce the dimensionality of the multidimensional historical pathological space tensor. The network architecture is configured with an input layer channel dimension constant of 3, a 3x3 two-dimensional convolution kernel size parameter, a stride constant of 1, and an edge filling mode parameter specifically set to an all-zero filling constant mode. The nonlinear activation function rectified linear unit operator is called, and a pre-specified hidden layer weight bias constant matrix is ​​loaded. Specifically, during the system compilation phase, the main thread reads the binary weight file from a fixed path on the local disk and maps it to the graphics processing unit's memory allocation address space. This drives the memory data pointer to slide and read historical sample data into the forward propagation network. The network channel calls the multidimensional array addressing operator to retrieve the corresponding coordinate nodes inside the matrix, extracts the pre-stored scalar data in the prediction cache register, constructs the expert diagnosis trajectory and prediction trajectory derivation phlegm-dampness error matrix generation scheme, extracts the pre-stored 36-dimensional expert diagnosis trajectory tensor and sets it as the benchmark target value, and calls the output prediction trajectory tensor of the near-end strategy optimization algorithm and sets it as the network prediction output scalar. It executes the spatial coordinate transformation instruction to align the benchmark target value and the network prediction output scalar, activates the floating-point subtractor core inside the arithmetic logic unit and issues the subtraction instruction, reads the aligned scalar memory address values ​​on both sides and executes the double-precision floating-point subtraction instruction to calculate the absolute difference, calls the absolute value calculation function to remove the negative sign flag, allocates a continuous memory block space constant of 1024 bytes, transports and writes the difference constant through the instruction data bus and collects the difference to construct the floating-point distribution space, and generates the phlegm-dampness conversion error array. The threshold filtering submodule, based on the phlegm-dampness conversion error array, uses a proportional-first empirical replay algorithm and a floating-point threshold comparison operator to divide the hierarchical error echelon boundary distribution. It sequentially reads the memory block to lock the array's memory starting address, drives the pointer to slide byte-by-byte to read array elements, defines the hierarchical parameters, configures the parameters for defining the hierarchical interval groups, sets the bottom line constant of 0 and the top line constant of 0.5 for the first interval, and sets the bottom line constant of 0.5 and the top line constant of 1.0 for the second interval. It compares element values ​​falling into the intervals and calls the probability density allocation function to calibrate the corresponding hierarchical probability floating-point values. It introduces a small positive offset parameter of 0.001 to execute addition instructions to avoid zero-probability value point anomalies, calls the power function operator to assign a priority level exponent constant of 0.6 to the summation value, and imports preset boundary constants. With a value of 0.8, the above-mentioned preset operation is as follows: during the device startup and loading phase, the main control program reads the fields of the underlying environment configuration file and converts them into floating-point constants, which are then fixed into read-only registers. The comparison processor core is activated to execute the instruction to compare the floating-point probability with the boundary constant. A Boolean classification mask matrix is ​​output. When the floating-point probability is greater than the boundary constant value of 0.8, the data block corresponding to the current pointer memory address is retained. Historical tensors higher than the boundary constant are extracted. A hard threshold truncation operator is executed to remove noise data blocks associated with the distribution probability of values ​​lower than the boundary constant. A lower limit constant of 95% is set for the high-risk pathological sample extraction hit rate, thereby providing a basis for the effectiveness of the filtering optimization technology. The data splicing function is called to package the out-of-bounds tensor data segment and push it to the experience cache pool stack to obtain the out-of-bounds physical condition intervention sample. The product-derivative merging submodule: Based on out-of-bounds physical intervention samples, it uses an importance sampling weight compensation algorithm and an adaptive moment estimation tensor update operator to extract the sampling multiplier coefficients within the samples. It calls the addressing optimization instruction set to locate the coordinates of the underlying samples and read the coefficient values. It configures the total capacity of the experience pool as an integer parameter of 10000, calls the reciprocal calculator to calculate the reciprocal constant of the capacity parameter, extracts the correlation probability parameter of out-of-bounds samples and calculates the reciprocal constant of the probability parameter, executes a double-precision multiplication instruction to multiply the reciprocal constants of two terms, calls the exponential operation function to introduce a compensation factor constant of 0.4 to calculate and generate independent sampling coefficient floating-point numbers, and issues floating-point multiplication to the arithmetic logic unit. The instruction multiplies the sampling coefficients and the error gradient constant. The preset error gradient constant value is 0.005. The above preset operation is as follows: the hyperparameter scheduling module reads the corresponding floating-point parameter value of the global variable data dictionary memory segment and pushes it onto the execution stack. The product value is calculated to form a penalty scalar. The preset fixed address block memory of the graphics processing unit is retrieved to store the network layer weight parameter array. The high-speed memory bus is connected to extract the pre-stored corresponding parameters. The tensor corresponding addition instruction is called to merge the penalty scalar and the pre-stored corresponding parameters. The numerical superposition operation is performed and the original weight parameter memory address block data is overwritten to establish an iteratively updated parameter tensor.

[0027] A convolutional neural network reads historical sample data of Qi and Blood constitution, constructs a spatial array of Qi and Blood constitution, sets local convolution kernel parameters, controls the local convolution kernel to perform sliding traversal in the spatial array of Qi and Blood constitution, calculates the product constant of the same position elements, summarizes the product constant to generate the latent pathological tensor, calls the nonlinear activation operator to map the latent pathological tensor, defines the dimension reduction interval and extracts the peak constant of the interval, reduces the size of the latent pathological tensor, and outputs the Qi and Blood pathological feature array.

[0028] The adversarial feedback derivation module includes: Trajectory comparison and classification submodule: Based on the action evaluation numerical matrix and iterative update parameter tensor, extract the pre-stored expert dialectical trajectory tensor, retrieve the corresponding predicted trajectory tensor and align the corresponding spatial coordinate elements, calculate the absolute value difference between the baseline target value and the network prediction output scalar, and obtain the dialectical offset scalar array. The target tensor derivative submodule: Based on the dialectical offset scalar array, it extracts the entropy parameters of the Qi and Blood information inside the array, merges the entropy parameters with the scalar array to form a joint target feedback tensor, alternately issues the element derivative increase and decrease instructions, stops the operation when the boundary scalar is reached, and obtains the instantaneous dynamic reward constant. The trajectory comparison and classification submodule, based on the action evaluation numerical matrix and iteratively updated parameter tensors, uses the tensor L1 norm absolute difference operator and multidimensional spatial coordinate alignment algorithm to extract pre-stored expert diagnostic trajectory tensors. It calls high-speed memory direct read instructions to lock specific sector blocks on the solid-state drive, reads the preset standard diagnostic timing array, and specifically, during the system initialization phase, the main process reads the hexadecimal encoded block inside the medical knowledge base configuration table and maps it to the static address pool of read-only memory. It then extracts the pre-packaged 36-dimensional expert trajectory tensor data block, retrieves the corresponding predicted trajectory tensor, and initiates cross-channel data transfer via the graphics processing unit's video memory bus. According to the block, extract the model forward propagation prediction coordinate array, call the coordinate axis matching function to execute the spatial dimension reduction reconstruction instruction, input the alignment dimension parameter constant 3, issue the tensor reshaping instruction to align the corresponding spatial coordinate elements, activate the floating-point subtractor core inside the arithmetic logic unit, read the two alignment tensor corresponding memory address values ​​to execute the double-precision floating-point subtraction instruction, call the absolute value calculator to remove the negative sign flag, calculate the difference between the baseline target value and the network prediction output scalar absolute value, import the tensor splicing function to set the splicing dimension index constant 0, loop through all floating-point constants in the three-dimensional feature space and merge to output the absolute difference constant set, and obtain the dialectical offset scalar array; The target tensor differentiation submodule: Based on the dialectical offset scalar array, it uses the Shannon information entropy calculation rule and tensor cascade splicing operator to extract the Qi and blood information entropy parameters within the array. It reads the floating-point probability of the Qi and blood distribution within the array, calls the logarithmic function calculator to execute a logarithmic calculation instruction with the natural constant as the base, issues a floating-point multiplication instruction to the arithmetic logic unit to multiply the probability floating-point number and the corresponding logarithmic value, calls the accumulator array to execute a summation instruction and flips the positive and negative sign bits to generate the Qi and blood information entropy scalar, allocates a continuous memory block space constant of 512 bytes, calls the tensor cascade splicing operator to set the cascade axis parameter to 1, merges the Qi and blood information entropy parameters and the scalar array to form a joint target feedback tensor, and uses a backpropagation optimization algorithm over time and a gradient truncation boundary control operator to alternately issue element derivative increase / decrease instructions. It imports a preset learning rate hyperparameter of 0.001. The above preset operations... Specifically, the compilation environment reads the hyperparameter tuning script fields, forcibly casts them to global constants, and writes them to registers. It then activates the automatic differentiation computation graph, traverses the nodes in reverse, and applies the chain rule to calculate the partial derivative constant. It issues a gradient descent instruction to subtract the learning rate multiplied by the gradient product from the current weight value. It calls the amplitude limiting control function to detect the absolute gradient value, compares the gradient value with the boundary threshold constant 0.5, and forcibly truncates and replaces the constant 0.5 if the value is greater than the boundary threshold constant 0.5. It stops the operation when the boundary scalar is reached, schedules the sparse matrix compression operator to perform continuous encoding of non-zero element coordinates, releases redundant all-zero matrix data blocks inside the graphics processing unit's memory, sets the peak constant limit of memory usage to 4096 megabytes, thereby providing support for the effect of memory space overhead compression technology, pushes the convergence constant to the state buffer, and obtains the instantaneous dynamic reward constant.

[0029] Please see Figure 1 The multidimensional dynamic identification and early warning method for body-disease-symptom includes the following steps: S1: Based on the clinical four diagnostic parameters and the scores of the constitution assessment scale, the constitution spatial architecture is mapped and the Yin deficiency tensor value is injected. The intersection of the phlegm-dampness internal resistance matrix is ​​extracted, the tensor inner product summation constant is calculated, the top-level source level coordinate starting point is defined, and the belief state particle set is obtained. S2: Based on the coordinates of the particle set of belief state, extract historical time series features and calculate the Qi and Blood tensor by dividing it by the frequency constant decay scalar and then dividing the boundary tensor. Backtrack the branch to accumulate the constitution flow frequency, sort the descending position to filter the first sequence, and output the action code of the first Jingfang. S3: Based on the action encoding of the first classical formula, the deep architecture of the constitution features is mapped and the tensors of spleen and stomach transportation and liver qi stagnation are separated. The correlation tensors are multiplied and added and the product values ​​are merged. A scaling factor is introduced to scale and truncate the tensors and sum the overflow scalars to generate the action evaluation numerical matrix. S4: Based on the numerical matrix of motion evaluation, traverse the spatial array of Qi and blood samples, align the expected scalar subtraction difference to form the phlegm and dampness error matrix, calibrate the hierarchical probability to extract the constant samples of the limit limit, multiply by the sampling coefficient and the error gradient constant, and establish the iterative update parameter tensor. S5: Based on the numerical matrix of motion evaluation and the iteratively updated parameter tensor, align the dialectical and simulated evolution trajectory, calculate the numerical difference to obtain the classification constant, merge the blood and qi parameters to splice the target tensor, alternately increase and decrease the derivative until the boundary scalar stops, and obtain the instantaneous dynamic reward constant.

[0030] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A multidimensional dynamic identification and early warning system for body-disease-symptom, characterized in that, The system includes: Latent pathogenesis optimization module: By using clinical four diagnostic parameters and physical constitution assessment scale scores, the physical constitution pathology array is defined, the Yin deficiency tensor and phlegm-dampness internal obstruction matrix are entered, the inner product summation value is calculated and defined as the top source level, and the belief state particle set is obtained. Dynamic Evolutionary Game Theory Module: Based on the coordinates of the particle set of the belief state, the constant is calculated by dividing the Qi and Blood tensor by the frequency through the Long Short-Term Memory Network, the boundary tensor is extracted and the constitution flow frequency is back-stacked, the first and second sequence is selected, and the action code of the first Jingfang is output. Competition flow splitting module: Based on the action encoding of the first prescription, extract the constitution tensor and split it into the spleen and stomach transportation and transformation branch and the liver qi stagnation branch. Calculate the output tensors of the two branches, add the tensor values ​​and perform scaling operation to generate the action evaluation numerical matrix. Priority playback optimization module: Based on the action evaluation numerical matrix, the predicted values ​​are subtracted by the convolutional neural network to form the phlegm and dampness error matrix, the probability floating-point numbers are calibrated and samples above the limit constant are extracted, the coefficients are multiplied by the gradient constant, and the iterative update parameter tensor is established. The adversarial feedback derivation module: Based on the action evaluation numerical matrix and the iterative update parameter tensor, it compares the dialectical and simulated trajectories to obtain the classification constant, merges the qi and blood to construct the target tensor, and alternately executes the derivative rise and fall until it drops to the boundary and stops, thus obtaining the instantaneous dynamic reward constant.

2. The multidimensional dynamic identification and early warning system for body-disease-symptom as described in claim 1, characterized in that, The belief state particle set includes a Yin deficiency tensor, a phlegm-dampness internal resistance matrix, and a source-level position scalar. The primary classical formula action encoding includes a boundary tensor, a constitution flow frequency value, and a core classical formula blocking intervention constant. The action evaluation numerical matrix includes a spleen and stomach transportation and transformation branch output tensor, a liver qi stagnation branch output tensor, and a tensor overflow scalar. The iterative update parameter tensor includes a probability floating-point number, a sampling multiplier coefficient, an error gradient constant, and a penalty scalar. The instantaneous dynamic reward constant includes a classification constant, a syndrome differentiation offset scalar array, and a qi and blood information entropy parameter.

3. The multidimensional dynamic identification and early warning system for body-disease-symptom as described in claim 1, characterized in that, The latent pathogenesis optimization module includes: The array structure construction submodule: Based on clinical characterization parameters and body mass scale scores, it constructs a three-dimensional coordinate system of body constitution, inputs the values ​​of the Yin deficiency and dryness-heat bias tensor and extracts the intersection points of the phlegm-dampness internal obstruction matrix, aligns multiple pathological element scalars, summarizes the multidimensional pathological data architecture, and generates a multidimensional bias feature matrix. Source-level partitioning submodule: Based on the multidimensional biased feature matrix, extract multiple row and column space coordinate elements, calculate the tensor element inner product summation constant, summarize the summed values ​​to define the initial boundary points, set the source-level position scalar, match the bottom architecture exploration starting point, and obtain the belief state particle set.

4. The multidimensional dynamic identification and early warning system for body-disease-symptom as described in claim 1, characterized in that, The dynamic evolution game module includes: Qi and Blood Tensor Sampling Submodule: Based on the coordinates of the belief state particle set, deploy a long short-term memory network to cyclically extract the evolution characteristics of Qi and Blood time series, divide the space of Qi and Blood circulation tensor array, read the frequency scalar and floating-point constant of Qi deficiency Yang micro, issue an instruction to divide the tensor elements by the frequency scalar, merge the floating-point constant to reconstruct the exploration data matrix, and obtain the evolution state exploration constant. Branch frequency accumulation submodule: Based on the evolution state exploration constant, compare the internal Qi and blood boundary values ​​of the constant, extract the phlegm-dampness stagnation tensor and trace the branch path to the corresponding level coordinates, extract the constitution flow frequency value, issue an addition command to accumulate the level-related frequency value, and obtain the state backtracking value array. First-position sequence selection submodule: Based on the state backtracking numerical array, read the spatial coordinates of the elements inside the array, compare and filter the numerical values ​​of the array elements, arrange the descending position sequence of the array elements, extract the action sequence associated with the first-position value and extract the core classical prescription blocking intervention constant, and obtain the action code of the first classical prescription.

5. The multidimensional dynamic identification and early warning system for body-disease-symptom as described in claim 1, characterized in that, The Long Short-Term Memory (LSTM) network extracts the Qi and Blood temporal tensor sequence, initializes the hidden state tensor and cell state constant, extracts the current Qi and Blood tensor within the Qi and Blood temporal tensor sequence, executes the inner product summation instruction of the hidden state tensor and the current Qi and Blood tensor to generate an intermediate state matrix, calls the activation operator to calculate the intermediate state matrix to obtain the retention probability matrix, simultaneously combines the current Qi and Blood tensor to execute the inner product instruction to generate candidate incremental tensors, multiplies the retention probability matrix and the cell state constant to obtain the forgotten state tensor, superimposes the forgotten state tensor and the candidate incremental tensor and overwrites the cell state constant, issues a tensor mapping instruction for the overwritten cell state constant, and outputs the Qi and Blood temporal evolution feature array.

6. The multidimensional dynamic identification and early warning system for body-disease-symptom as described in claim 1, characterized in that, The competitive flow splitting module includes: Tensor Cutting and Splitting Submodule: Based on the action encoding of the first prescription, map the feature parameters of the constitution tensor dimension, split the spleen and stomach transportation and transformation and liver qi stagnation branches, extract the tensor related to spleen and stomach and liver qi stagnation, multiply the tensor related to spleen and stomach and liver qi stagnation and then merge the tensor product constant to obtain the split branch output array. Tensor numerical scaling submodule: Based on the split branch output array, extract independent tensor data within the array, add the values ​​of the corresponding elements of the independent tensors, introduce a proportional constant to scale the tensor sum value, define the parameter mapping boundary to truncate the tensor overflow scalar, and generate an action evaluation numerical matrix.

7. The multidimensional dynamic identification and early warning system for body-disease-symptom as described in claim 1, characterized in that, The priority playback optimization module includes: Bias Feature Construction Submodule: Based on the action evaluation numerical matrix, deploy a convolutional neural network pathological feature parsing architecture to read historical sample data of Qi and Blood constitution, retrieve the corresponding coordinate nodes inside the matrix, align the expected value with the actual predicted scalar, issue a subtraction instruction to calculate the absolute difference, collect the difference to construct a floating-point distribution space, and generate a phlegm-dampness transformation error array. The over-threshold screening submodule: Based on the phlegm-dampness transformation error array, it divides the hierarchical error echelon boundary distribution, reads the array elements to define the hierarchical parameters and marks the corresponding hierarchical probability floating-point values, compares the probability floating-point numbers with the boundary constant, extracts the historical tensors that are higher than the boundary constant, and obtains the over-limit constitution intervention samples. The product derivative merging submodule: Based on the over-boundary physical intervention samples, extract the sampling multiplier coefficients inside the samples and locate the coordinates of the bottom sample to read the coefficient values. Multiply the sampling coefficients with the error gradient constant, calculate the product value to form a penalty scalar, merge the penalty scalar with the pre-stored co-position parameters, and establish an iteratively updated parameter tensor.

8. The multidimensional dynamic identification and early warning system for body-disease-symptom as described in claim 1, characterized in that, The convolutional neural network reads historical sample data of Qi and Blood constitution, constructs a spatial array of Qi and Blood constitution, sets local convolution kernel parameters, controls the local convolution kernel to perform sliding traversal in the spatial array of Qi and Blood constitution, calculates the product constant of the same element, summarizes the product constant to generate the hidden pathological tensor, calls the nonlinear activation operator to map the hidden pathological tensor, defines the dimensionality reduction interval and extracts the peak constant of the interval, reduces the size of the hidden pathological tensor, and outputs the Qi and Blood pathological feature array.

9. The multidimensional dynamic identification and early warning system for body-disease-symptom as described in claim 1, characterized in that, The adversarial feedback derivation module includes: Trajectory comparison and classification submodule: Based on the action evaluation numerical matrix and the iterative update parameter tensor, extract the pre-stored expert dialectical trajectory tensor, retrieve the corresponding simulated evolution trajectory tensor and align the corresponding spatial coordinate elements, calculate the absolute numerical difference of the real scalar, and obtain the dialectical offset scalar array. The target tensor differentiation submodule extracts the entropy parameters of the Qi and Blood information within the array based on the dialectical offset scalar array, merges the entropy parameters with the scalar array to form a joint target feedback tensor, alternately issues element derivative increase and decrease commands, stops the operation when the boundary scalar is reached, and obtains the instantaneous dynamic reward constant.

10. A multidimensional dynamic identification and early warning method for body-disease-symptom, characterized in that, The execution of the multidimensional dynamic identification and early warning system for body-disease-symptom as described in any one of claims 1-9 includes the following steps: S1: Based on the clinical four diagnostic parameters and the scores of the constitution assessment scale, the constitution spatial architecture is mapped and the Yin deficiency tensor value is injected. The intersection of the phlegm-dampness internal resistance matrix is ​​extracted, the tensor inner product summation constant is calculated, the top-level source level coordinate starting point is defined, and the belief state particle set is obtained. S2: Based on the coordinates of the particle set of the belief state, extract the historical time sequence features and calculate the Qi and Blood tensor divided by the frequency constant decay scalar to divide the boundary tensor. Backtrack the branch to accumulate the constitution flow frequency, sort the descending position to filter the first sequence, and output the first Jingfang action code. S3: Based on the action encoding of the first prescription, map the deep architecture constitution features and separate the tensors of spleen and stomach transportation and liver qi stagnation, multiply and add the correlation tensors and merge the product values, introduce a scaling coefficient to scale and truncate the tensors and sum the overflow scalars to generate the action evaluation numerical matrix. S4: Based on the action evaluation numerical matrix, traverse the Qi and Blood sample spatial array, align the expected scalar subtraction difference to form the phlegm and dampness error matrix, calibrate the hierarchical probability to extract the out-of-limit constant sample, multiply by the sampling coefficient and the error gradient constant, and establish the iterative update parameter tensor. S5: Based on the action evaluation numerical matrix and the iterative update parameter tensor, align the dialectical and simulated evolution trajectories, calculate the numerical difference to obtain the classification constant, merge the blood and qi parameters to splice the target tensor, alternately increase and decrease the derivative until the boundary scalar stops, and obtain the instantaneous dynamic reward constant.

Images