Traditional Chinese medicine syndrome differentiation intelligent analysis method and system based on measurement data of meridian instrument

Abstract

The invention provides a traditional Chinese medicine syndrome differentiation intelligent analysis method and system based on meridian instrument measurement data, and relates to the technical field of medical data analysis. Preprocessing the measurement data sample of the meridian instrument; through a convolutional neural network and discrete wavelet transform, feature extraction is carried out on the preprocessed meridian instrument measurement data samples, and fusion is carried out to obtain a combined feature vector; constructing an intelligent analysis model; inputting the combined feature vector into an intelligent analysis model for analysis to obtain a predicted value of the meridian state; with the purpose of minimizing the error between the predicted value and the true value of the meridian state, a traditional Chinese medicine five-element theory is fused, and a comprehensive loss function is constructed; training the model through a back propagation algorithm and an optimizer according to the comprehensive loss function; acquiring real-time meridian instrument measurement data; and analyzing the real-time measurement data of the meridian instrument through the trained intelligent analysis model to obtain a real-time predicted value of the meridian state.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data analysis, and in particular to a method and system for intelligent analysis of TCM syndrome differentiation based on meridian instrument measurement data. Background Art

[0002] With the continuous development of TCM diagnostic technology, especially the widespread application of meridian instruments, TCM practitioners can use the data measured by meridian instruments to more objectively analyze the health status of patients. As an important tool for TCM diagnosis, meridian instruments provide quantitative physiological data by measuring the resistance changes of human meridians, which helps to assist TCM practitioners in more accurate syndrome differentiation analysis. However, how to interpret these data efficiently and accurately has become a key challenge to improving the quality of diagnosis.

[0003] In the existing technology, the traditional manual interpretation method mainly relies on experienced Chinese medicine practitioners to make subjective judgments and diagnoses based on the data measured by the meridian instrument. Although this method can provide personalized diagnosis based on the patient's overall condition, it is highly dependent on the experience of Chinese medicine practitioners, resulting in low efficiency, lack of unified standards for diagnostic results, and difficulty meeting the requirements of modern medicine for objectivity and efficiency.

[0004] In addition, judging the state of meridians by setting fixed thresholds lacks scientific basis and cannot handle complex pattern recognition problems, resulting in low diagnostic accuracy and failure to fully reflect the patient's health status. Summary of the invention

[0005] In order to solve the technical problems that the traditional TCM syndrome differentiation intelligent analysis method based on meridian instrument measurement data is highly dependent on the experience of TCM practitioners, resulting in low efficiency, lack of unified standards for diagnosis results, and difficulty in meeting the requirements of modern medicine for objectivity and efficiency, and that the meridian status is judged by setting fixed thresholds, lacks scientific basis, and cannot handle complex pattern recognition problems, resulting in low diagnostic accuracy and failure to fully reflect the patient's health status, the present invention provides a TCM syndrome differentiation intelligent analysis method and system based on meridian instrument measurement data.

[0006] The technical solution provided by the embodiment of the present invention is as follows:

[0007] First aspect:

[0008] An embodiment of the present invention provides a TCM syndrome differentiation intelligent analysis method based on meridian instrument measurement data, comprising:

[0009] S1: Obtaining the meridian instrument measurement data sample;

[0010] S2: preprocessing the meridian instrument measurement data sample;

[0011] S3: Through convolutional neural network and discrete wavelet transform, feature extraction is performed on the preprocessed meridian instrument measurement data samples respectively, and the combined feature vector is obtained by fusion;

[0012] S4: Construct an intelligent analysis model based on the attention mechanism of bidirectional long short-term memory network;

[0013] S5: inputting the combined feature vector into the intelligent analysis model for analysis to obtain a predicted value of the meridian state;

[0014] S6: With the goal of minimizing the error between the predicted value and the true value of the meridian state, integrating the theory of the five elements of traditional Chinese medicine, constructing a comprehensive loss function of the intelligent analysis model;

[0015] S7: According to the comprehensive loss function, the intelligent analysis model is trained through a back propagation algorithm and an optimizer;

[0016] S8: obtaining real-time meridian instrument measurement data;

[0017] S9: Analyze the real-time meridian instrument measurement data through the trained intelligent analysis model to obtain a real-time prediction value of the meridian state.

[0018] Second aspect:

[0019] An embodiment of the present invention provides a TCM syndrome differentiation intelligent analysis system based on meridian instrument measurement data, comprising: a memory and one or more processors;

[0020] One or more application programs are stored in the memory, and the one or more application programs are suitable for being executed by the one or more processors to implement the above-mentioned intelligent analysis method of TCM syndrome differentiation based on the measurement data of the meridian instrument.

[0021] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0022] In the present invention, by constructing an intelligent analysis model of a bidirectional long short-term memory network based on an attention mechanism, it is no longer highly dependent on the experience of traditional Chinese medicine practitioners. Through the analysis of the intelligent analysis model, the predicted value of the meridian state is obtained with high efficiency, so that the diagnosis results have a unified standard, and can meet the requirements of modern medicine for objectivity and efficiency. By minimizing the error between the predicted value and the true value of the meridian state as the goal, integrating the theory of the five elements of traditional Chinese medicine, and constructing a comprehensive loss function of the intelligent analysis model, according to the comprehensive loss function, the intelligent analysis model is trained through the back propagation algorithm and the optimizer, avoiding the judgment of the meridian state by setting a fixed threshold. It has a scientific basis and can effectively deal with complex pattern recognition problems, thereby greatly improving the diagnostic accuracy and comprehensively reflecting the health status of the patient. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0024] Figure 1 A schematic diagram of a flow chart of a method for intelligent analysis of TCM syndrome differentiation based on meridian instrument measurement data provided by an embodiment of the present invention;

[0025] Figure 2 A schematic diagram of the structure of a TCM syndrome differentiation intelligent analysis system based on meridian instrument measurement data provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0026] The technical solution of the present invention is described below in conjunction with the accompanying drawings.

[0027] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to indicate examples, illustrations or explanations. Any embodiment or design described as "example" in the present invention should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of the word "example" is intended to present the concept in a specific way. In addition, in the embodiments of the present invention, the meaning expressed by "and / or" can be both, or it can be either of the two.

[0028] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0029] Reference Manual Attached Figure 1 , shows a flow chart of a method for intelligent analysis of TCM syndrome differentiation based on meridian instrument measurement data provided by an embodiment of the present invention.

[0030] The embodiment of the present invention provides a method for intelligent analysis of TCM syndrome differentiation based on meridian instrument measurement data, which can be implemented by a TCM syndrome differentiation intelligent analysis device based on meridian instrument measurement data, which can be a terminal or a server. The processing flow of the method for intelligent analysis of TCM syndrome differentiation based on meridian instrument measurement data may include the following steps:

[0031] S1: Obtain the meridian instrument measurement data sample.

[0032] Specifically, the meridians of multiple patients are measured using a meridian instrument, and resistance data of 24 meridian points (original acupoints) are collected to form meridian instrument measurement data samples.

[0033] In the present invention, by collecting data from 24 meridian points, the meridian status of the human body can be fully reflected. These data contain rich information, which can provide a broad basis for subsequent feature extraction and analysis, and enhance the comprehensiveness and accuracy of diagnosis.

[0034] S2: Preprocess the meridian instrument measurement data samples.

[0035] In the present invention, by preprocessing the meridian instrument data acquired in real time, interference in the data can be quickly removed to ensure that the model can make accurate real-time predictions based on high-quality data, which is crucial to the stability and reliability of the real-time diagnosis system.

[0036] In a possible implementation, S2 specifically includes sub-steps S201 to S203:

[0037] S201: Cleaning the meridian instrument measurement data samples.

[0038] In the present invention, data cleaning can ensure that the data distribution of different patients is consistent and reduce fluctuations caused by individual differences or measurement errors. This helps the model to focus more on common patterns during training and avoid overfitting certain patient-specific noise data, thereby improving the model's generalization ability for new data.

[0039] S202: De-noising the meridian instrument measurement data samples after data cleaning by using wavelet transform technology.

[0040] Among them, wavelet transform technology is a mathematical tool for signal processing. It can provide time domain and frequency domain information of the signal at the same time by decomposing the signal into components of different frequencies and time scales. It performs local analysis on the signal by using short-time oscillating wavelet functions. It is suitable for capturing mutations, anomalies and complex time-frequency characteristics in the signal. Therefore, it is widely used in image processing, data compression, medical signal analysis and other fields.

[0041] In the present invention, the meridian instrument measurement process may be affected by environmental factors, measurement equipment noise or individual differences of patients, resulting in unnecessary noise in the measurement data. Wavelet transform is a powerful signal processing tool that can analyze data at multiple scales and accurately remove noise. Compared with traditional filtering methods, wavelet transform can better retain useful information in the signal and remove high-frequency noise.

[0042] In a possible implementation, S202 specifically includes sub-steps S2021 to S2025:

[0043] S2021: Perform discrete wavelet transform on the meridian instrument measurement data sample after data cleaning to obtain the first wavelet coefficient:

[0044]

[0045] Among them, DWT i (a,b) represents the first wavelet coefficient obtained by discrete wavelet transform of the meridian instrument measurement data sample after data cleaning, a represents the scale parameter, b represents the translation parameter, f represents the translation parameter, i (t) represents the i-th meridian instrument measurement data sample after data cleaning at time t, and ψ represents the wavelet function.

[0046] Among them, discrete wavelet transform (DWT) is a signal processing technology used to decompose signals into different frequency components and time scales. Unlike continuous wavelet transform, discrete wavelet transform calculates by sampling discrete points of the signal and uses discrete wavelet basis functions to analyze the time-frequency characteristics of the signal. DWT can effectively capture local features, mutations and irregularities in the signal, and is particularly suitable for processing non-stationary signals. It is often used in applications such as image compression, denoising, and biomedical signal analysis, and can provide details and overall information of the signal in different time periods.

[0047] S2022: Estimate the noise standard deviation based on the first wavelet coefficient:

[0048]

[0049] Among them, σ i represents the noise standard deviation of the meridian instrument measurement data sample after data cleaning, median() represents the median, d i,1 It represents the first wavelet coefficient of the meridian instrument measurement data sample at the finest scale after the i-th data cleaning.

[0050] S2023: Calculate the denoising threshold based on the noise standard deviation:

[0051]

[0052] Among them, λ i represents the denoising threshold of the meridian instrument measurement data sample after the i-th data cleaning, N i Represents the length of the i-th meridian instrument measurement data sample after data cleaning.

[0053] S2024: Perform soft threshold processing on the first wavelet coefficient according to the denoising threshold:

[0054] η(DWT i (a,b))=sign(DWT i (a,b))·max(|DWT i (a,b)|-λ i ,0)

[0055] Among them, η(DWT i (a,b)) represents the first wavelet coefficient after soft threshold processing, sign() represents the sign function, and max() represents the maximum value.

[0056] S2025: Perform inverse discrete wavelet transform on the first wavelet coefficient after soft threshold processing to obtain the meridian instrument measurement data sample after denoising processing:

[0057] f i ′(t)=IDWT(η(DWT i (a,b)))

[0058] Among them, f i ′(t) represents the i-th meridian instrument measurement data sample after denoising, and IDWT() represents inverse discrete wavelet transform.

[0059] Among them, the Inverse Discrete Wavelet Transform (IDWT) is a process used to reconstruct the original signal after the signal is decomposed by the Discrete Wavelet Transform (DWT). Through the IDWT, the wavelet coefficients at different frequencies and time scales after decomposition are recombined to restore a version that is as close to the original signal as possible. This process uses the low-frequency and high-frequency components generated during wavelet decomposition and inversely applies the wavelet basis functions to complete the reconstruction of the signal. IDWT is widely used in applications such as image reconstruction and signal denoising to ensure that the original structure of the signal can be restored after analysis and processing.

[0060] In the present invention, through multi-scale decomposition, signals in different frequency bands can be analyzed separately and noise can be removed in a targeted manner. High-frequency components usually contain noise, while low-frequency components retain the overall trend of the signal, which can be processed more accurately. Adaptive noise estimation reduces the error accumulation problem caused by artificially setting parameters, so that the model can more accurately process the noise of different patients and different data sets, thereby improving the overall data quality. The calculated threshold can balance denoising and signal retention, avoid signal loss caused by excessive denoising, and thus retain more information useful for diagnosis. Soft threshold processing can smoothly reduce the impact of noise, make the denoised signal smoother and more natural, improve the overall effect of denoising processing, and avoid signal mutations that may be caused by hard thresholds. By inverse transforming the denoised wavelet coefficients, a cleaner original signal can be restored, unnecessary noise can be removed, and important information required for diagnosis can be retained.

[0061] S203: Standardizing the meridian instrument measurement data samples after the denoising process.

[0062] In the present invention, the normalization process enables the model to better generalize to unseen test data or new patient data, avoiding the model from overfitting to a specific training data distribution. In particular, when processing data from different patients, normalization allows the model to identify the common features of the data without being affected by the individual numerical differences.

[0063] S3: Through convolutional neural network and discrete wavelet transform, feature extraction is performed on the preprocessed meridian instrument measurement data samples respectively, and the combined feature vector is obtained by fusion.

[0064] Among them, convolutional neural network (CNN) is a deep learning model designed for processing structured data such as images. It can automatically extract features and perform classification or regression tasks. CNN extracts local features and combines them layer by layer by applying convolutional layers, pooling layers, and fully connected layers to the input data to identify patterns and structures in the data. The convolutional layer uses convolution kernels (filters) to scan the input data and capture local spatial relationships. The pooling layer reduces the data dimension by downsampling and retains important information. CNN is widely used in image recognition, object detection, natural language processing and other fields, and is favored for its automatic feature extraction capabilities and advantages in processing high-dimensional data.

[0065] In the present invention, the time domain feature vector extracted from the convolutional neural network and the frequency domain feature vector extracted from the discrete wavelet transform are fused to generate a combined feature vector that contains the time domain dynamic information and frequency domain frequency information of the data. This fusion helps the model to make full use of all aspects of the data when making decisions, increasing the accuracy of the prediction.

[0066] In a possible implementation, S3 specifically includes sub-steps S301 to S303:

[0067] S301: extracting the time domain features of the preprocessed meridian instrument measurement data samples through a convolutional neural network to obtain a time domain feature vector.

[0068] In the present invention, the convolutional neural network (CNN) can automatically extract key time domain features from the data, avoiding the tedious process of manually designing features. Traditional methods require manual feature selection, while CNN can automatically extract effective time series patterns from large amounts of data, reducing the uncertainty caused by human subjective judgment.

[0069] Among them, the convolutional neural network includes multiple one-dimensional convolutional layers, pooling layers and fully connected layers.

[0070] In a possible implementation, S301 specifically includes sub-steps S3011 to S3013:

[0071] S3011: In the one-dimensional convolution layer of the convolutional neural network, extract the local time domain features of the preprocessed meridian instrument measurement data samples:

[0072] Y l =f(W l *X (l-1) +b l )

[0073] Among them, Y l represents the local time domain features output by the l-th one-dimensional convolutional layer, f() represents the activation function, and W l represents the convolution kernel of the lth one-dimensional convolution layer, * represents the convolution operation, X (l-1) represents the local time domain features output by the l-1th one-dimensional convolutional layer, b l It represents the bias term of the l-th one-dimensional convolutional layer, l = 1, 2, …, L, and L represents the total number of one-dimensional convolutional layers.

[0074] Optionally, the activation function is a ReLU activation function.

[0075] It should be noted that for the local time domain features output by the first one-dimensional convolutional layer, X 0 =X.

[0076] S3012: In the pooling layer of the convolutional neural network, a maximum pooling operation is performed on the local time domain features:

[0077]

[0078] in, Represents the local time domain features after the maximum pooling operation output by the i-th pooling window, max() means taking the maximum value, represents the local time domain features of the input i-th pooling window, s represents the step size, and k represents the size of the pooling kernel.

[0079] S3013: In the fully connected layer of the convolutional neural network, the local time domain features after the maximum pooling operation are integrated to obtain the time domain feature vector:

[0080] T=Z=W fc ·F′+b fc

[0081] F′=flatten(P L )

[0082] Among them, Z represents the time domain feature vector, W fc represents the weight of the fully connected layer, F' represents the feature vector after the flattening operation, and b fc represents the bias of the fully connected layer, flatten() represents the flattening operation, P LIt represents the local time domain features output by the Lth one-dimensional convolutional layer, that is, the local time domain features output by the last one-dimensional convolutional layer.

[0083] In the present invention, through the local perception of the convolution layer and the dimensionality reduction of the pooling layer, the model can process large-scale data with less computing resources, ensuring high efficiency in real-time applications. For example, the amount of data of the meridian instrument may be large, but through the operation of gradual dimensionality reduction, the scale of the features can be significantly reduced, and the reasoning speed and training efficiency of the model can be improved. The fully connected layer can combine the local time domain features into a global feature vector by flattening the pooled features and performing linear combinations. This operation enables the model to combine the global information of the signal, rather than relying solely on local features. This global feature vector is critical for the model to accurately predict the state of the meridians, and can help the model comprehensively analyze the relationship between different local features.

[0084] S302: extracting the frequency domain features of the pre-processed meridian instrument measurement data samples through discrete wavelet transform to obtain a frequency domain feature vector.

[0085] In the present invention, discrete wavelet transform can separate noise from signal when decomposing the signal. The high-frequency part usually contains more noise information, while the low-frequency part reflects the main characteristics of the signal. By extracting the low-frequency approximate coefficients, the model can focus more on the core features of the data, reduce noise interference, and improve the accuracy of diagnosis. The frequency domain features generated by discrete wavelet transform are signal features captured from both the overall and local aspects. This multi-dimensional information helps the model to perform more robustly when processing new data.

[0086] In a possible implementation, S302 specifically includes sub-steps S3021 to S3024:

[0087] S3021: Perform discrete wavelet transform on the pre-processed meridian instrument measurement data samples to obtain the second wavelet coefficients:

[0088]

[0089] Among them, DWT i ′(a,b) represents the second wavelet coefficient obtained by discrete wavelet transform of the i-th preprocessed meridian instrument measurement data sample, a represents the scale parameter, b represents the translation parameter, and f i (t) represents the i-th preprocessed meridian instrument measurement data sample at time t, and ψ represents the wavelet function.

[0090] S3022: Decompose the second wavelet coefficients according to a preset number of decomposition layers to obtain approximate coefficients and detail coefficients of each layer.

[0091] It should be noted that those skilled in the art can set the size of the decomposition layer number according to actual needs, and the present invention is not limited here.

[0092] S3023: Extract multiple frequency domain features of approximate coefficients and detail coefficients of each layer:

[0093] μ j =mean(cA j ) or mean(cD j )

[0094] σ j =std(cA j ) or std(cD j )

[0095] or

[0096] H j =-∑(p j ·log2(p j ))

[0097] Among them, μ j represents the mean of the approximate coefficient or detail coefficient of the jth layer, cA j represents the approximation coefficient of the jth layer, cD j represents the detail coefficient of the jth layer, σ j represents the standard deviation of the approximation coefficient or detail coefficient of the jth layer, E j Represents the energy of the approximate coefficient or detail coefficient of the jth layer, H j represents the entropy of the approximate coefficient or detail coefficient of the jth layer, p j Represents the probability distribution of the approximate coefficient or detail coefficient of the j-th layer, j=1,2,…,L', L' represents the total number of preset decomposition levels.

[0098] S3024: Combine each frequency domain feature to obtain a frequency domain feature vector:

[0099] F=[μ1,σ1,E1,H1,…,μ L′ ,σ L′ ,E L′ ,H L′ ]

[0100] Among them, μ1 represents the mean of the approximate coefficient or detail coefficient of the first layer, σ1 represents the standard deviation of the approximate coefficient or detail coefficient of the first layer, E1 represents the energy of the approximate coefficient or detail coefficient of the first layer, H1 represents the entropy of the approximate coefficient or detail coefficient of the first layer, μ L′ Represents the mean of the approximate coefficient or detail coefficient of the L'th layer, σ L′Indicates the standard deviation of the approximation coefficient or detail coefficient of the L'th layer, E L′ Represents the energy of the approximate coefficient or detail coefficient of the L'th layer, H L′ Represents the entropy of the approximation coefficient or detail coefficient of the L'th layer.

[0101] In the present invention, the meridian instrument measurement data can be decomposed into different frequency components through discrete wavelet transform. The decomposed signal is represented by approximate coefficients and detail coefficients, which respectively reflect the low-frequency global characteristics and high-frequency local characteristics of the signal. The number of decomposition layers is flexibly set to ensure that signal features can be captured from different frequency ranges. Through different decomposition layers, the model can extract frequency domain information of different granularities as needed. By extracting statistical features such as mean, standard deviation, energy and entropy, the model can capture the overall trend, variation amplitude, signal strength and complexity of the signal in the frequency domain. By extracting features from the approximate coefficients and detail coefficients of multiple decomposition layers and combining them, a comprehensive frequency domain feature vector can be generated. This feature combination enhances the richness of feature representation and provides more comprehensive information for the model, enabling it to identify complex health patterns and anomalies.

[0102] S303: Fusing the time domain feature vector with the frequency domain feature vector to obtain a combined feature vector:

[0103] X combined =[T; F]

[0104] Among them, X combined represents the combined eigenvector, T represents the time domain eigenvector, and F represents the frequency domain eigenvector.

[0105] In this paper, time domain features can capture the changes of signals over time, while frequency domain features reveal the periodicity and frequency components in the signal. Fusion of time domain and frequency domain features can help the model fully understand the data from different dimensions and avoid information loss caused by relying on only a single type of feature.

[0106] Optionally, use principal component analysis to reduce the feature dimensionality of the combined feature vector:

[0107] Compute the covariance matrix of the combined eigenvectors:

[0108] ∑=cov(X combined )

[0109] Among them, Σ represents the covariance matrix, and cov() represents the function of calculating the covariance matrix.

[0110] Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and eigenvectors, and sort them in descending order:

[0111] ∑v′=λ0v′

[0112] Among them, v' represents the eigenvector and λ0 represents the eigenvalue.

[0113] According to the size of the eigenvalue, select the first k eigenvectors and construct the eigenvector matrix:

[0114] V k =[v1′,v2′,…,v′ k ]

[0115] Among them, V k represents the eigenvector matrix, [v1′,v2′,…,v′ k ] represents the first k eigenvectors.

[0116] Project the combined eigenvector into the eigenvector matrix to obtain the combined eigenvector after reducing the feature dimension:

[0117] X final =X combined ·V k

[0118] Among them, X final Represents the combined feature vector after reducing the feature dimension.

[0119] Furthermore, in the subsequent steps, the combined feature vector after reducing the feature dimension through the principal component analysis method is input into the intelligent analysis model for analysis to obtain the predicted value of the meridian state.

[0120] Among them, principal component analysis (PCA) is a statistical technique for data dimensionality reduction. It extracts the most representative information by projecting high-dimensional data into a lower-dimensional space. PCA calculates a set of orthogonal principal components by finding the main direction of change in the data. Each principal component is a linear combination of the data and is sorted by variance. The first few principal components can usually explain most of the variance in the data, thereby retaining important features and discarding noise or redundant information. PCA is widely used in data compression, noise reduction, and visualization to help simplify complex data and improve the computational efficiency of the model.

[0121] In the present invention, by reducing the dimension through the principal component analysis method, the dimension of the eigenvector can be greatly reduced, thereby reducing the computational complexity of the model and improving the running speed of the model. This is particularly important for processing large-scale meridian data or real-time analysis, and can ensure that the system can make predictions quickly under limited computing resources. By selecting the first k eigenvectors with the largest eigenvalues, it is possible to ensure that the most important information in the signal is retained, while discarding those minor features that contribute less to the total variation. In this way, the eigenvectors input to the model will be more compact and effective, reducing unnecessary calculations and improving the sensitivity of the model to the main features.

[0122] S4: Construct an intelligent analysis model based on the bidirectional long short-term memory network with attention mechanism.

[0123] Among them, the bidirectional long short-term memory network (BiLSTM) is a recursive neural network based on the long short-term memory network (LSTM), which can simultaneously process forward and reverse information in sequence data. BiLSTM uses two LSTM layers, one to process data from front to back (forward LSTM) and the other to process data from back to front (reverse LSTM), and then combines these two parts of information for prediction. This structure can capture long-term dependencies in the sequence, and when processing tasks such as natural language processing and time series analysis, it can use past and future contextual information to significantly improve the accuracy and performance of predictions.

[0124] In the present invention, an intelligent analysis model based on a bidirectional long short-term memory network with an attention mechanism is constructed, which can fully capture the temporal dependencies in meridian data, flexibly process complex health signals, and adaptively focus on the most important information. The attention mechanism enables the model to dynamically adjust its focus on features, improving the diagnostic accuracy, interpretability and generalization ability of the model. The bidirectional long short-term memory network can process past and future information simultaneously, enhancing the understanding of complex health patterns, making the model more practical and robust in TCM diagnosis.

[0125] S5: Input the combined feature vector into the intelligent analysis model for analysis to obtain the predicted value of the meridian state.

[0126] In the present invention, by inputting the fused combined feature vector into the model, which contains both time domain features and frequency domain features, the model can analyze meridian data from multiple angles. This multi-dimensional feature enables the model to understand the meridian state from two levels: time change and frequency distribution, thereby improving the accuracy and comprehensiveness of the prediction.

[0127] Among them, the intelligent analysis model includes input layer, bidirectional long short-term memory network layer, attention layer, fully connected layer and output layer.

[0128] In a possible implementation, S5 specifically includes sub-steps S501 to S505:

[0129] S501: In the input layer of the intelligent analysis model, a combined feature vector is input.

[0130] Optionally, the intelligent analysis model also includes: an embedding layer.

[0131] In the embedding layer of the intelligent analysis model, the combined feature vector is mapped to a higher dimensional representation space.

[0132] In the present invention, the embedding layer can map the input combined feature vector to a higher-dimensional representation space, expanding the expressive power of the features. This high-dimensional representation can capture more complex patterns and relationships in the original features, especially for the multi-dimensional features (time domain and frequency domain features) in the meridian instrument measurement data, the embedding layer can help the model better extract the potential relationships of these features.

[0133] S502: In the bidirectional long short-term memory network layer of the intelligent analysis model, the combined feature vector is processed to generate a forward hidden state and a reverse hidden state for each time step.

[0134] Optionally, the number of hidden units in the bidirectional long short-term memory network layer of the smart analysis model is set to 128 or 256.

[0135] In the present invention, the bidirectional LSTM generates forward and reverse hidden states to ensure that the model can use past features and refer to future features, so as to more accurately predict the current meridian state. In particular, when detecting meridian change trends, the bidirectional structure can better analyze the directionality and periodicity of these changes.

[0136] Among them, the bidirectional long short-term memory network layer includes a forward long short-term memory network layer and a reverse long short-term memory network layer.

[0137] In a possible implementation, S502 specifically includes sub-steps S5021 and S5022:

[0138] S5021: Generate the forward hidden state of each time step through the forward long short-term memory network layer:

[0139] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0140] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0141]

[0142] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0143] h t =ot ⊙tanh(c t )

[0144] Among them, f t represents the forget gate of the forward LSTM network layer, W f represents the weight matrix of the forget gate of the forward LSTM network layer, h t-1 represents the hidden state of the t-1th time step of the forward LSTM network layer, x t represents the input of the tth time step, b f represents the bias term of the forget gate of the forward LSTM network layer, i t represents the input gate of the forward LSTM network layer, W i represents the weight matrix of the input gate of the forward LSTM network layer, b i represents the bias term of the input gate of the forward LSTM network layer, represents the candidate meridian state of the forward long short-term memory network layer, W c The weight matrix representing the candidate meridian states of the forward LSTM network layer, b c represents the bias term of the candidate meridian state of the forward long short-term memory network layer, c t represents the meridian state of the current time step of the forward long short-term memory network layer, c t-1 represents the meridian state at the t-1th time step of the forward long short-term memory network layer, o t Represents the output gate of the forward LSTM layer.

[0145] S5022: Generate the reverse hidden state of each time step through the reverse long short-term memory network layer:

[0146]

[0147] in, represents the forget gate of the reverse LSTM network layer, represents the weight matrix of the forget gate of the reverse LSTM network layer, represents the hidden state of the t+1th time step of the reverse long short-term memory network layer, represents the bias term of the forget gate of the reverse LSTM network layer, represents the input gate of the reverse LSTM network layer, represents the weight matrix of the input gate of the reverse LSTM network layer, represents the bias term of the input gate of the reverse LSTM network layer, represents the candidate meridian states of the reverse long short-term memory network layer, The weight matrix representing the candidate meridian states of the reverse long short-term memory network layer, The bias term of the candidate meridian state of the reverse long short-term memory network layer, c′ t represents the meridian state of the current time step of the reverse long short-term memory network layer, represents the meridian state of the reverse long short-term memory network layer at the t+1th time step, represents the output gate of the reverse LSTM network layer, represents the weight matrix of the output gate of the reverse LSTM network layer, Represents the bias term of the output gate of the reverse LSTM network layer.

[0148] In the present invention, through the gating mechanism of LSTM (forget gate, input gate, output gate), the forward LSTM can effectively retain important historical information while discarding irrelevant noise data. This enables the model to accurately grasp key changes in the time series, especially in the long-term change analysis of meridian data, the model can better capture the changing trend of meridian status. The characteristic of the reverse LSTM is that it can obtain information from future time steps, which is particularly useful when analyzing multi-level meridian signals. For example, in the current prediction, future changes may contain key information, and the model can correct the judgment of the current state based on future time steps, thereby improving the sensitivity to complex signals and the accuracy of prediction.

[0149] S503: In the attention layer of the intelligent analysis model, the context vector is calculated based on the forward hidden state and the reverse hidden state through the Bahdanau attention mechanism:

[0150]

[0151] α t =softmax(e t )

[0152]

[0153] Among them, c′ t represents the context vector, α t represents the attention weight, ⊙ represents element-wise multiplication, and h t represents the forward hidden state, represents the reverse hidden state, softmax() represents the softmax function, e t represents the attention score, v represents the weight vector of the attention score, T represents the transpose operation, tanh() represents the hyperbolic tangent activation function, W a The weight matrix representing the attention score, b a Represents the bias term of the attention score.

[0154] Specifically, through the Bahdanau attention mechanism, in the attention layer of the intelligent analysis model, the attention score e is calculated for each time step t , and use the softmax function to convert the attention score e t Converted to the corresponding attention weight α t , combining the forward hidden state and the reverse hidden state, through the attention weight α t Perform a weighted summation on each time step to generate a weighted context vector c′ t .

[0155] Among them, the Bahdanau attention mechanism, also known as additive attention, is an attention mechanism used in sequence-to-sequence models (such as machine translation, text generation, etc.). It helps the model better focus on the input part that is most relevant to the current output by dynamically calculating the importance weight of each input sequence element during the decoding process. Specifically, Bahdanau attention calculates an attention score for each input element, assigns weights based on these scores, and generates a context vector to enhance the model's ability to capture long-distance dependency information. This mechanism makes the model more efficient when processing long sequences and improves the ability to understand complex tasks.

[0156] In the present invention, through the Bahdanau attention mechanism, the model can assign a specific weight to each time step, and these weights can be used to explain the decision basis of the model. Doctors can analyze these attention weights to understand which time steps contribute the most to the prediction of meridian status, thereby improving their understanding and trust in the diagnosis results.

[0157] S504: In the fully connected layer of the intelligent analysis model, the context vector is input to generate the predicted value of the meridian state:

[0158] y t =σ(W o ·c′ t +b o )

[0159] Among them, y t represents the predicted value of the meridian state, σ represents the sigmoid function, W o represents the weight matrix of the output gate of the forward LSTM network layer, b o Represents the bias term of the output gate of the forward LSTM network layer.

[0160] It should be noted that the context vector c′ is processed in the fully connected layer of the intelligent analysis model. t , which has integrated the information of the forward LSTM network layer and the reverse LSTM network layer, namely the forward hidden state h tand the reverse hidden state Therefore, only a set of weight matrices and bias terms are needed to generate the final predicted value of the meridian state. This set of parameters (weight matrices and bias terms) is not specific to the forward LSTM network layer or the reverse LSTM network layer, but acts on the entire context vector c′ t .

[0161] For example, the weight matrix and bias term can be the weight matrix W of the output gate of the forward long short-term memory network layer o and the bias term b o It can also be the weight matrix of the output gate of the reverse long short-term memory network layer and the bias term Therefore, only one set of parameters is needed (the weight matrix W of the output gate of the forward LSTM network layer o and the bias term b o Or the weight matrix of the output gate of the reverse LSTM network layer and the bias term ) to generate the final predicted value of the meridian state. When the weight matrix W of the output gate of the forward long short-term memory network layer is applied o and the bias term b o As shown in sub-step S504. Similarly, when the weight matrix of the output gate of the reverse long short-term memory network layer is applied and the bias term hour:

[0162]

[0163] Among them, y t represents the predicted value of the meridian state, σ represents the sigmoid function, represents the weight matrix of the output gate of the reverse LSTM network layer, Represents the bias term of the output gate of the reverse LSTM network layer.

[0164] Optionally, one or more fully connected layers are added to the intelligent analysis model, and the number of neurons in each layer (such as 128 neurons) can be set by oneself, indicating that this layer has 128 outputs.

[0165] In the present invention, the context vector is a feature representation after the attention mechanism combines the forward and reverse hidden states, which contains rich global information. Through the fully connected layer, the model can further integrate the context vector to generate a comprehensive representation of the meridian state. This step helps the model integrate complex time series and frequency domain features into a simplified, easy-to-interpret global feature, providing strong support for prediction tasks.

[0166] S505: In the output layer of the intelligent analysis model, the predicted value of the meridian state is output.

[0167] Optionally, the output layer of the intelligent analysis model depends on the specific task:

[0168] When the task is binary classification (for example, predicting healthy or unhealthy), the number of neurons in the output layer is 1, and the Sigmoid activation function is used to map the output value to the interval [0,1].

[0169] When the task is multi-classification (for example, classifying different types of meridian states), the number of output neurons is the number of categories, and the Softmax function is used to normalize the output to the probability of each category.

[0170] In the present invention, by flexibly selecting the output layer structure, the model has strong scalability. When it is necessary to process different classification tasks, it is easy to adapt by simply adjusting the number of neurons and activation functions in the output layer. For diagnostic tasks that may be added in the future (such as introducing more categories of meridian states), the model only needs to expand the output layer and keep the structure of other layers unchanged. This flexibility ensures the long-term adaptability of the model.

[0171] It should be noted that for the meridian diagnosis task, the output of all time steps can be averaged or the output of the last time step can be used as the final prediction, t = 1, 2, ..., T:

[0172] y final =mean(y1,y2,…,y T ) or y final =y T

[0173] Among them, y final represents the final predicted value of the meridian state, mean() represents the average function, (y1,y2,…,y T ) represents the predicted value of the meridian state output at each time step, y T represents the predicted value of the meridian state output at the last time step, and T represents the total number of time steps.

[0174] In the present invention, the model can flexibly adapt to time series data of different lengths by using the average output or the output of the last time step. This makes the model more flexible and adaptable when processing different types of meridian data. Whether it is measurement data of a short period of time or health monitoring data of a long period of time, the model can make correct judgments through appropriate output strategies.

[0175] S6: With the goal of minimizing the error between the predicted value and the true value of the meridian status, the comprehensive loss function of the intelligent analysis model is constructed by integrating the Five Elements Theory of Traditional Chinese Medicine.

[0176] Among them, the Five Elements Theory of Traditional Chinese Medicine is one of the basic theories of traditional Chinese medicine. It corresponds the five basic substances or energies in nature - wood, fire, earth, metal, and water - to the five internal organs of the human body (liver, heart, spleen, lung, and kidney), and explains the balance and health of the human body through the relationship of mutual generation and mutual restraint. The mutual generation (mutual promotion) and mutual restraint (mutual restriction) between the five elements embody the principle of dynamic balance and determine the coordinated operation of various organs, meridians, emotions, etc. in the human body. The Five Elements Theory is used to guide TCM diagnosis and treatment, and maintain or restore the health of the human body by regulating the balance of the five elements.

[0177] In the present invention, by incorporating the Five Elements theory into the loss function, the model can learn the mutual relationship and dynamic balance between the Five Elements, thereby more accurately predicting the health status of different meridians and improving the accuracy of prediction. In this way, the model no longer relies solely on mathematical error minimization, but can also combine medical knowledge to generate prediction results that are more in line with actual physiological laws.

[0178] In a possible implementation, S6 specifically includes sub-steps S601 and S602:

[0179] S601: With the goal of minimizing the error between the predicted value and the true value of the meridian state, a cross entropy loss function is constructed:

[0180] L loss =-∑(y true ·log(y pred ))

[0181] Among them, L loss represents the cross entropy loss function, y true Indicates the true value of the meridian state, y pred Represents the predicted value of the meridian status.

[0182] S602: Integrate the theory of the five elements of traditional Chinese medicine into the cross entropy loss function to construct a comprehensive loss function for the intelligent analysis model:

[0183] L total =L loss +λ′·L wuxing

[0184] L wuxing =∑(w mn ·|y m -y n |)

[0185] Among them, L total represents the comprehensive loss function, λ' represents the weight coefficient of the regularization term of the Five Elements Theory of Traditional Chinese Medicine, L wuxing represents the regularization term of the Five Elements Theory of Traditional Chinese Medicine, wmn represents the five-element relationship weight between meridian m and meridian n, y m represents the predicted value of the meridian state corresponding to meridian m, y n Represents the predicted value of the meridian state corresponding to meridian n.

[0186] In the present invention, through the cross entropy loss function, the model can accurately quantify the error between the predicted value and the true value, thereby optimizing the model's prediction performance and improving the accuracy of meridian diagnosis. By integrating the Five Elements Theory, the model can better reflect the interaction between different meridians, help capture the overall health status, and improve the comprehensiveness and accuracy of the diagnostic results. By integrating the Five Elements Theory into the loss function, the model not only relies on modern scientific algorithms, but is also based on the classical theories of traditional Chinese medicine, which enhances the interpretability of the diagnostic results.

[0187] S7: According to the comprehensive loss function, the intelligent analysis model is trained through the back propagation algorithm and optimizer.

[0188] Among them, the back propagation algorithm is an optimization algorithm for training neural networks. It calculates the error between the model's predicted value and the true value, and propagates the error from the output layer to the input layer layer by layer to adjust the weight and bias of each layer. Back propagation calculates the partial derivatives of the loss function with respect to the parameters of each layer through the gradient descent method, updates the network parameters, and makes the model gradually approach the optimal solution. It is the core of neural network training and can efficiently optimize multi-layer neural networks, helping the model to improve prediction accuracy and learn complex patterns.

[0189] Optionally, the optimizer is an Adam optimizer.

[0190] Among them, the Adam optimizer is an adaptive optimization algorithm for training neural networks, combining the advantages of momentum gradient descent and RMSProp. Adam dynamically adjusts the learning rate of each parameter by simultaneously calculating the first-order moment (moving average of the gradient) and the second-order moment (moving average of the square of the gradient), so that it can converge faster and maintain stability during the training process. It can adaptively adjust the learning rate and is suitable for processing large-scale data and high-dimensional parameters. Due to its high efficiency and low memory requirements, it is widely used in the optimization of deep learning models.

[0191] In the present invention, the back propagation algorithm can efficiently optimize the model parameters, improve the diagnostic accuracy of the model by minimizing the comprehensive loss function, and is suitable for large-scale data processing. The Adam optimizer can accelerate the convergence speed of the model, improve the training efficiency, and maintain stability during the training process, which is suitable for processing complex multi-dimensional meridian data.

[0192] S8: Obtain real-time meridian instrument measurement data.

[0193] S9: Analyze the real-time meridian instrument measurement data through the trained intelligent analysis model to obtain the real-time prediction value of the meridian status.

[0194] In the present invention, by inputting the real-time data measured by the meridian instrument into the model, the system can instantly analyze and generate prediction results. Users can understand the current meridian status at any time, which provides timely feedback for health monitoring and helps users quickly respond to early signs of health problems. By analyzing the user's real-time meridian data, the model can perform personalized analysis based on the user's individual characteristics and historical health data. For example, the meridian status baseline of different users may be different, and the model can adjust the prediction results based on individual differences to provide personalized health management solutions for each user.

[0195] Furthermore, the real-time prediction value of the meridian status is visualized through the SHAP value analysis method:

[0196] Construct a weighted linear model and calculate the SHAP value of each feature:

[0197]

[0198] Among them, φ i represents the SHAP value of feature i, f represents the intelligent analysis model, x represents the real-time meridian instrument measurement data, S represents the subset that does not contain feature i, N represents the set of all features, and f represents the real-time meridian instrument measurement data. x (S∪{i}) represents the model output for prediction using the time steps and feature i of subset S, f x (S) represents the model output using only the subset S of time steps for prediction.

[0199] Among them, feature i refers to a specific meridian point (original acupoint) in the real-time meridian measurement data or its corresponding specific time step data. For example, for each meridian point (original acupoint), there may be multiple time steps of data collection, and feature i can be the resistance, potential and other measurement values ​​of a certain original acupoint at a certain time step. It represents a dimension or a specific attribute in the input data used for the prediction model.

[0200] It should be noted that the SHAP value analysis method (Shapley Additive Explanations) is a technique for explaining the output of machine learning models. It is based on the Shapley value in cooperative game theory and can quantify the contribution of each feature to the model's prediction results. SHAP calculates the impact of a feature on the prediction results when it is added to the model and evaluates the average contribution of the feature in all possible combinations, thereby providing a unified way to explain the impact of each input feature of the model on the final prediction. SHAP values ​​have global and local explanatory capabilities, and can clearly show the impact of features on a single prediction result or the entire model behavior, enhancing the interpretability and transparency of the model.

[0201] It should be noted that since it is computationally infeasible to calculate all possible combinations of subsets S, the KernelSHAP algorithm can be used to approximate the SHAP value:

[0202] Combine the real-time meridian meter measurement data with the background dataset to generate synthetic samples:

[0203] z′=x′·z

[0204] Among them, z' represents the synthetic sample, x' represents the binary mask of the real-time meridian meter measurement data, and z represents the background dataset.

[0205] Calculate the kernel weights for the synthetic samples:

[0206]

[0207] Among them, π x′ (z′) represents the kernel weight of the synthetic sample, M represents the number of features, C represents the combination number formula, and |z′| represents the number of selected features.

[0208] Based on the kernel weights, a weighted linear model is trained to calculate the approximate SHAP value of each feature:

[0209]

[0210] Among them, φ represents the approximate SHAP value, Z represents the set of all synthetic samples, f(z') represents the model output using synthetic samples for prediction, and φ0 represents the bias term of the intelligent analysis model.

[0211] Specifically, for each feature in the meridian diagnosis model, we calculate its SHAP value:

[0212] for i in range(num_features):

[0213] shap_value_i=calculate_shap(model,feature_i,instance)

[0214] Furthermore, the SHAP values ​​can be visualized using the following method:

[0215] Use summary graphs to show the impact of SHAP values ​​on different features:

[0216] plot_summary=lambda shap_values:

[0217] sort(abs(mean(|shap_values|,axis=0)))

[0218] Generate a dependency graph showing how the SHAP value of a feature changes with its value:

[0219] plot_dependence=lambda feature,shap_values:

[0220] scatter(feature_values,shap_values)

[0221] The generated image attempts to display the prediction results of a single real-time meridian meter measurement data, clearly showing the positive or negative contribution of each feature to the prediction result:

[0222] plot_force=lambda shap_values:

[0223] sum(base_value+shap_values)

[0224] Furthermore, the features are sorted according to the absolute value of the SHAP value to obtain the importance ranking of the global features:

[0225] importance=sort(mean(|shap_values|, axis=0), descending=True)

[0226] Furthermore, for a single predicted real-time meridian instrument measurement data, the contribution of each feature to the final result is generated:

[0227] local_explanation=base_value+shap_values

[0228] Furthermore, the SHAP values ​​of multiple real-time meridian instrument measurement data are summarized to generate a global feature importance explanation:

[0229] global_importance=mean(|shap_values|, axis=0)

[0230] Furthermore, the interaction SHAP value analysis is used to further explain the impact of the interaction between features on model prediction:

[0231] interaction_values=shap.TreeExplainer(model).shap_interaction_values(X)

[0232] Furthermore, a portion of samples are randomly selected from the training data as the background data set, and the SHAP value is calculated for each predicted real-time meridian instrument measurement data:

[0233] for instance in test_data:

[0234] shap_values=explainer.shap_values(instance)

[0235] Visualizing SHAP values:

[0236] shap.summary_plot(shap_values,test_data,feature_names=feature_names)

[0237] shap.dependence_plot("feature_name",shap_values,test_data)

[0238] shap.force_plot(explainer.expected_value,shap_values[0],test_data[0])

[0239] Analyzing feature importance:

[0240] feature_importance=np.abs(shap_values).mean(0)

[0241] importance_df=pd.DataFrame(list(zip(feature_names,feature_importance)),columns=['feature','importance'])

[0242] importance_df=importance_df.sort_values('importance', ascending=False)

[0243] Interpretation of individual diagnostic results:

[0244] for i,feature in enumerate(feature_names):

[0245] print(f"{feature}:{shap_values[0][i]}").

[0246] It should be noted that the KernelSHAP algorithm is an efficient method for approximating SHAP values, and is suitable for complex machine learning models where Shapley values ​​cannot be directly calculated. KernelSHAP samples different subsets of input features and uses a weighted linear model based on a kernel function to estimate the contribution of each feature to model prediction. Since it is computationally infeasible to directly calculate all possible feature combinations, KernelSHAP greatly reduces the computational complexity through reasonable sampling and weighting methods. It can quickly calculate the SHAP value of each feature while maintaining the interpretability of the results, and is widely used to improve model interpretability.

[0247] In the present invention, the SHAP value provides a way to assign the contribution of each feature to the model prediction, so that we can clearly see which features (such as resistance and potential values ​​at a certain time step) have a positive contribution to the prediction of the meridian state and which features have a negative impact. This explanation can not only help users understand the output of the model, but also enhance trust in the diagnosis results. Through the force plot, the positive or negative contribution of each feature to the final prediction result can be intuitively displayed, and the user or doctor can understand at a glance how the model reaches the current diagnosis. This is of great help to the transparency and trust of the diagnostic process.

[0248] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0249] In the present invention, by constructing an intelligent analysis model of a bidirectional long short-term memory network based on an attention mechanism, it is no longer highly dependent on the experience of traditional Chinese medicine practitioners. Through the analysis of the intelligent analysis model, the predicted value of the meridian state is obtained with high efficiency, so that the diagnosis results have a unified standard, and can meet the requirements of modern medicine for objectivity and efficiency. By minimizing the error between the predicted value and the true value of the meridian state as the goal, integrating the theory of the five elements of traditional Chinese medicine, and constructing a comprehensive loss function of the intelligent analysis model, according to the comprehensive loss function, the intelligent analysis model is trained through the back propagation algorithm and the optimizer, avoiding the judgment of the meridian state by setting a fixed threshold. It has a scientific basis and can effectively deal with complex pattern recognition problems, thereby greatly improving the diagnostic accuracy and comprehensively reflecting the health status of the patient.

[0250] Reference Manual Attached Figure 2 , showing a schematic diagram of the structure of a TCM syndrome differentiation intelligent analysis system based on meridian instrument measurement data provided by the present invention.

[0251] The present invention further provides a TCM syndrome differentiation intelligent analysis system 30 based on the meridian instrument measurement data, comprising: a memory 303 and one or more processors 301 .

[0252] The memory 303 stores one or more application programs, and the one or more application programs are suitable for being executed by the one or more processors 301 to implement the TCM syndrome differentiation intelligent analysis method based on meridian instrument measurement data described in the method embodiment.

[0253] The TCM syndrome differentiation intelligent analysis system 30 based on the meridian instrument measurement data includes: a processor 301 and a memory 303. The processor 301 and the memory 303 are connected, such as through a bus 302.

[0254] The structure of the intelligent TCM syndrome differentiation analysis system 30 based on the meridian instrument measurement data does not constitute a limitation on the embodiment of the present invention.

[0255] Processor 301 may be a CPU, a general purpose processor, a DSP, an ASIC, an FPGA or other programmable logic device, a transistor logic device, a hardware component or any combination thereof. It may implement or execute various exemplary logic blocks, modules and circuits described in conjunction with the disclosure of the present invention. Processor 301 may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0256] The bus 302 may include a path to transmit information between the above components. The bus 302 may be a PCI bus or an EISA bus, etc. The bus 302 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.

[0257] The memory 303 can be a ROM or other type of static storage device that can store static information and instructions, a RAM or other type of dynamic storage device that can store information and instructions, or an EEPROM, a CD-ROM or other optical disk storage, an optical disk storage (including a compressed optical disk, a laser disk, an optical disk, a digital versatile disk, a Blu-ray disk, etc.), a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited to these.

[0258] It should be noted that the TCM syndrome differentiation intelligent analysis system 30 based on meridian instrument measurement data can implement the above-mentioned TCM syndrome differentiation intelligent analysis method based on meridian instrument measurement data, and can achieve the same or similar technical effects. To avoid repetition, the present invention will not go into details.

[0259] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0260] In the present invention, by constructing an intelligent analysis model of a bidirectional long short-term memory network based on an attention mechanism, it is no longer highly dependent on the experience of traditional Chinese medicine practitioners. Through the analysis of the intelligent analysis model, the predicted value of the meridian state is obtained with high efficiency, so that the diagnosis results have a unified standard, and can meet the requirements of modern medicine for objectivity and efficiency. By minimizing the error between the predicted value and the true value of the meridian state as the goal, integrating the theory of the five elements of traditional Chinese medicine, and constructing a comprehensive loss function of the intelligent analysis model, according to the comprehensive loss function, the intelligent analysis model is trained through the back propagation algorithm and the optimizer, avoiding the judgment of the meridian state by setting a fixed threshold. It has a scientific basis and can effectively deal with complex pattern recognition problems, thereby greatly improving the diagnostic accuracy and comprehensively reflecting the health status of the patient.

[0261] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which can be loaded and executed by a processor according to the intelligent analysis method for TCM syndrome differentiation based on the meridian instrument measurement data described in the first aspect.

[0262] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

[0263] There are a few points to note:

[0264] (1) The drawings of the embodiments of the present invention only relate to the structures related to the embodiments of the present invention, and other structures may refer to the general design.

[0265] (2) For the sake of clarity, in the drawings used to describe the embodiments of the present invention, the thickness of the layers or regions is exaggerated or reduced, that is, these drawings are not drawn according to the actual scale. It is understood that when an element such as a layer, film, region or substrate is referred to as being "on" or "under" another element, the element may be "directly" "on" or "under" the other element or there may be intermediate elements.

[0266] (3) In the absence of conflict, the embodiments of the present invention and the features therein may be combined with each other to obtain new embodiments.

[0267] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. The protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for intelligent analysis of TCM syndrome differentiation based on meridian instrument measurement data, characterized in that: include: S1: Obtaining the meridian instrument measurement data sample; S2: preprocessing the meridian instrument measurement data sample; S3: Through convolutional neural network and discrete wavelet transform, feature extraction is performed on the preprocessed meridian instrument measurement data samples respectively, and the combined feature vector is obtained by fusion; S4: Construct an intelligent analysis model based on the attention mechanism of bidirectional long short-term memory network; S5: inputting the combined feature vector into the intelligent analysis model for analysis to obtain a predicted value of the meridian state; S6: With the goal of minimizing the error between the predicted value and the true value of the meridian state, integrating the theory of the five elements of traditional Chinese medicine, constructing a comprehensive loss function of the intelligent analysis model; S7: According to the comprehensive loss function, the intelligent analysis model is trained through a back propagation algorithm and an optimizer; S8: obtaining real-time meridian instrument measurement data; S9: Analyze the real-time meridian instrument measurement data through the trained intelligent analysis model to obtain a real-time prediction value of the meridian state.

2. The intelligent analysis method for TCM syndrome differentiation based on meridian instrument measurement data according to claim 1, characterized in that: The S2 specifically includes: S201: performing data cleaning on the meridian instrument measurement data sample; S202: performing denoising on the meridian instrument measurement data samples after data cleaning by using wavelet transform technology; S203: Standardizing the meridian instrument measurement data samples after the denoising process.

3. The intelligent analysis method for TCM syndrome differentiation based on meridian instrument measurement data according to claim 2, characterized in that: The S202 specifically includes: S2021: Perform discrete wavelet transform on the meridian instrument measurement data sample after data cleaning to obtain the first wavelet coefficient: Among them, DWT i (a, b) represents the first wavelet coefficient obtained by discrete wavelet transform of the meridian instrument measurement data sample after data cleaning, a represents the scale parameter, b represents the translation parameter, f represents the i (t) represents the i-th meridian instrument measurement data sample after data cleaning at time t, ψ represents the wavelet function; S2022: Estimate the noise standard deviation according to the first wavelet coefficient: Among them, σ i represents the noise standard deviation of the meridian instrument measurement data sample after data cleaning, median() represents the median, d i,1 represents the first wavelet coefficient of the meridian instrument measurement data sample at the finest scale after the i-th data cleaning; S2023: Calculate a denoising threshold according to the noise standard deviation: Among them, λ i represents the denoising threshold of the meridian instrument measurement data sample after the i-th data cleaning, N i represents the length of the meridian instrument measurement data sample after the i-th data cleaning; S2024: Perform soft threshold processing on the first wavelet coefficient according to the denoising threshold: The(DWT i (a,b))=sign(DWT i (a,b))·max(|DWT i (a,b)|-λ i ,0) Among them, η(DWT i (a, b)) represents the first wavelet coefficient after soft threshold processing, sign() represents the sign function, and max() represents the maximum value; S2025: Perform inverse discrete wavelet transform on the first wavelet coefficient after soft threshold processing to obtain the meridian instrument measurement data sample after denoising processing: in i ′(t)=IDWT(η(DWT i (a,b))) Among them, f i ′(t) represents the i-th meridian instrument measurement data sample after denoising, and IDWT() represents inverse discrete wavelet transform.

4. The intelligent analysis method for TCM syndrome differentiation based on meridian instrument measurement data according to claim 1, characterized in that: The S3 specifically includes: S301: extracting the time domain features of the preprocessed meridian instrument measurement data samples through a convolutional neural network to obtain a time domain feature vector; S302: extracting the frequency domain features of the pre-processed meridian instrument measurement data samples through discrete wavelet transform to obtain a frequency domain feature vector; S303: Fusing the time domain feature vector with the frequency domain feature vector to obtain a combined feature vector: X combined =[T;F] Among them, X combined represents the combined eigenvector, T represents the time domain eigenvector, and F represents the frequency domain eigenvector.

5. The intelligent analysis method for TCM syndrome differentiation based on meridian instrument measurement data according to claim 4, characterized in that: The convolutional neural network includes a plurality of one-dimensional convolutional layers, a pooling layer and a fully connected layer, and S301 specifically includes: S3011: extracting the local time domain features of the preprocessed meridian instrument measurement data samples in the one-dimensional convolution layer of the convolutional neural network: Y l =f(W l *X (l-1) +b l ) Among them, Y l represents the local time domain features output by the l-th one-dimensional convolutional layer, f() represents the activation function, and W l represents the convolution kernel of the lth one-dimensional convolution layer, * represents the convolution operation, X (l-1) represents the local time domain features output by the l-1th one-dimensional convolutional layer, b l represents the bias term of the l-th one-dimensional convolutional layer, l = 1, 2, ..., L, L represents the total number of one-dimensional convolutional layers; S3012: In the pooling layer of the convolutional neural network, a maximum pooling operation is performed on the local time domain features: in, Represents the local time domain features after the maximum pooling operation output by the i-th pooling window, max() means taking the maximum value, represents the local time domain features of the input i-th pooling window, s represents the step size, and k represents the size of the pooling kernel; S3013: In the fully connected layer of the convolutional neural network, the local time domain features after the maximum pooling operation are integrated to obtain a time domain feature vector: T=Z=W fc ·F′+b fc F′=flatten(P L ) Among them, Z represents the time domain feature vector, W fc represents the weight of the fully connected layer, F' represents the feature vector after the flattening operation, and b fc represents the bias of the fully connected layer, flatten() represents the flattening operation, P L It represents the local time domain features output by the Lth one-dimensional convolutional layer, that is, the local time domain features output by the last one-dimensional convolutional layer.

6. The intelligent analysis method for TCM syndrome differentiation based on meridian instrument measurement data according to claim 4, characterized in that: The S302 specifically includes: S3021: Perform discrete wavelet transform on the pre-processed meridian instrument measurement data samples to obtain the second wavelet coefficients: Among them, DWT′ i (a, b) represents the second wavelet coefficient obtained by discrete wavelet transform of the i-th preprocessed meridian instrument measurement data sample, a represents the scale parameter, b represents the translation parameter, f represents the i (t) represents the i-th preprocessed meridian instrument measurement data sample at time t, ψ represents the wavelet function; S3022: Decomposing the second wavelet coefficients according to a preset number of decomposition layers to obtain approximate coefficients and detail coefficients of each layer; S3023: Extract multiple frequency domain features of approximate coefficients and detail coefficients of each layer: μ j =mean(cA j ) ormean(cD j ) σ j = std(cA j ) or std(cD j ) or H j =-∑(p j ·log2(p j )) Among them, μ j represents the mean of the approximate coefficient or detail coefficient of the jth layer, cA j represents the approximation coefficient of the jth layer, cD j represents the detail coefficient of the jth layer, σ j represents the standard deviation of the approximation coefficient or detail coefficient of the jth layer, E j Represents the energy of the approximate coefficient or detail coefficient of the jth layer, H j represents the entropy of the approximate coefficient or detail coefficient of the jth layer, p j represents the probability distribution of the approximate coefficient or detail coefficient of the j-th layer, j = 1, 2, ..., L', L' represents the total number of preset decomposition layers; S3024: Combine each frequency domain feature to obtain a frequency domain feature vector: F=[μ1,σ1,E1,H1,…,μ L′ ,s L′ ,E L′ ,H L′ ] Among them, μ1 represents the mean of the approximate coefficient or detail coefficient of the first layer, σ1 represents the standard deviation of the approximate coefficient or detail coefficient of the first layer, E1 represents the energy of the approximate coefficient or detail coefficient of the first layer, H1 represents the entropy of the approximate coefficient or detail coefficient of the first layer, μ L′ Represents the mean of the approximate coefficient or detail coefficient of the L'th layer, σ L′ Represents the standard deviation of the approximation coefficient or detail coefficient of the L'th layer, E L′ Represents the energy of the approximate coefficient or detail coefficient of the L'th layer, H L′ Represents the entropy of the approximation coefficient or detail coefficient of the L'th layer.

7. The intelligent analysis method for TCM syndrome differentiation based on meridian instrument measurement data according to claim 1, characterized in that: The intelligent analysis model includes an input layer, a bidirectional long short-term memory network layer, an attention layer, a fully connected layer and an output layer. S5 specifically includes: S501: In the input layer of the intelligent analysis model, input the combined feature vector; S502: In the bidirectional long short-term memory network layer of the intelligent analysis model, the combined feature vector is processed to generate a forward hidden state and a reverse hidden state for each time step; S503: In the attention layer of the intelligent analysis model, a context vector is calculated according to the forward hidden state and the reverse hidden state through the Bahdanau attention mechanism: a t =softmax(e t ) Among them, c′ t represents the context vector, α t represents the attention weight, ⊙ represents element-wise multiplication, and h t represents the forward hidden state, represents the reverse hidden state, softmax() represents the softmax function, e t represents the attention score, v represents the weight vector of the attention score, T represents the transpose operation, tanh() represents the hyperbolic tangent activation function, W a The weight matrix representing the attention score, b a represents the bias term of the attention score; S504: In the fully connected layer of the intelligent analysis model, the context vector is input to generate a predicted value of the meridian state: y t =σ(W o ·c′ t +b o ) Among them, y t represents the predicted value of the meridian state, σ represents the sigmoid function, W o represents the weight matrix of the output gate of the forward LSTM network layer, b o represents the bias term of the output gate of the forward LSTM network layer; S505: In the output layer of the intelligent analysis model, the predicted value of the meridian state is output.

8. The intelligent analysis method for TCM syndrome differentiation based on meridian instrument measurement data according to claim 7, characterized in that: The bidirectional long short-term memory network layer includes a forward long short-term memory network layer and a reverse long short-term memory network layer, and S502 specifically includes: S5021: Generate the forward hidden state of each time step through the forward long short-term memory network layer: f t =σ(W f ·[h t-1 ,x t ]+b f ) i t =σ(W i ·[h t-1 ,x t ]+b i ) the t =σ(W o ·[h t-1 ,x t ]+b o ) h t =o t ⊙tanh(c t ) Among them, f t represents the forget gate of the forward LSTM network layer, W f represents the weight matrix of the forget gate of the forward LSTM network layer, h t-1 represents the hidden state of the t-1th time step of the forward LSTM network layer, x t represents the input of the tth time step, b f represents the bias term of the forget gate of the forward LSTM network layer, i t represents the input gate of the forward LSTM network layer, W i represents the weight matrix of the input gate of the forward LSTM network layer, b i represents the bias term of the input gate of the forward LSTM network layer, represents the candidate meridian state of the forward long short-term memory network layer, W c The weight matrix representing the candidate meridian states of the forward LSTM network layer, b c represents the bias term of the candidate meridian state of the forward long short-term memory network layer, c t represents the meridian state of the current time step of the forward long short-term memory network layer, c t-1 represents the meridian state at the t-1th time step of the forward long short-term memory network layer, o t Represents the output gate of the forward long short-term memory network layer; S5022: Generate a reverse hidden state for each time step through the reverse long short-term memory network layer: in, represents the forget gate of the reverse LSTM network layer, represents the weight matrix of the forget gate of the reverse LSTM network layer, represents the hidden state of the reverse LSTM network layer at the t+1th time step, represents the bias term of the forget gate of the reverse LSTM network layer, represents the input gate of the reverse LSTM network layer, represents the weight matrix of the input gate of the reverse LSTM network layer, represents the bias term of the input gate of the reverse LSTM network layer, represents the candidate meridian states of the reverse long short-term memory network layer, The weight matrix representing the candidate meridian states of the reverse long short-term memory network layer, The bias term of the candidate meridian state of the reverse long short-term memory network layer, c′ t represents the meridian state of the current time step of the reverse long short-term memory network layer, represents the meridian state of the t+1th time step of the reverse long short-term memory network layer, represents the output gate of the reverse LSTM network layer, represents the weight matrix of the output gate of the reverse LSTM network layer, Represents the bias term of the output gate of the reverse LSTM network layer.

9. The intelligent analysis method for TCM syndrome differentiation based on meridian instrument measurement data according to claim 1, characterized in that: The S6 specifically includes: S601: With the goal of minimizing the error between the predicted value and the true value of the meridian state, a cross entropy loss function is constructed: L loss =-∑(and true ·log(y pred )) Among them, L loss represents the cross entropy loss function, y true Indicates the true value of the meridian state, y pred Indicates the predicted value of meridian status; S602: Integrate the theory of the five elements of traditional Chinese medicine into the cross entropy loss function to construct a comprehensive loss function of the intelligent analysis model: L total =L loss +λ′·L wuxing L wuxing =∑(w mn ·|y m -y n |) Among them, L total represents the comprehensive loss function, λ' represents the weight coefficient of the regularization term of the Five Elements Theory of Traditional Chinese Medicine, L wuxing represents the regularization term of the Five Elements Theory of Traditional Chinese Medicine, w mn represents the five-element relationship weight between meridian m and meridian n, y m represents the predicted value of the meridian state corresponding to meridian m, y n Represents the predicted value of the meridian state corresponding to meridian n.

10. A TCM syndrome differentiation intelligent analysis system based on meridian instrument measurement data, characterized in that: include: memory and one or more processors; One or more application programs are stored in the memory, and the one or more application programs are suitable for being executed by the one or more processors to implement the intelligent analysis method of TCM syndrome differentiation based on meridian instrument measurement data as described in any one of claims 1 to 9.

Images