Traditional Chinese medicine pulse condition diagnosis method based on encoder-decoder framework

Abstract

The invention relates to a traditional Chinese medicine pulse condition diagnosis method based on an encoder-decoder framework. The traditional Chinese medicine pulse condition diagnosis method comprises the steps of collecting data, preprocessing a data set, constructing a neural network based on the encoder-decoder framework, training the network and testing. According to the method, a pulse condition signal diagnosis problem is modeled into a machine translation problem, mapping is established between the pulse condition signal and the diagnosis text, the problem how to extract key features from the complex pulse condition signal by using the deep neural network is solved, the accuracy of pulse condition signal diagnosis is improved, and the diagnosis text for the pulse condition signal is directly generated.

Description

Technical Field

[0001] The invention relates to a Chinese medicine pulse diagnosis method based on an encoder-decoder framework, belonging to the technical field of medical diagnosis. Background Art

[0002] As an important part of traditional Chinese medicine, pulse diagnosis plays the role of a human information diagnosis system. This technology reflects the material and energy metabolism and functional regulation of the human body through pulse information. Since the publication of the Pulse Classic, the earliest pulse monograph in my country in the third century AD, the pulse theory has been continuously enriched and improved, which has had a profound impact on the global medical field. As a non-invasive detection technology, pulse diagnosis has been widely recognized and paid attention to by experts at home and abroad. With the in-depth integration of mathematics, information engineering, computer science and other disciplines in the field of traditional Chinese medicine, especially the application of deep learning technology, the learning and classification diagnosis capabilities of pulse signals have been greatly improved. This change has not only promoted the modernization of traditional Chinese medicine diagnosis, but also injected new vitality into the dissemination and development of traditional Chinese medicine around the world.

[0003] Nevertheless, the accuracy of TCM pulse judgment depends largely on the personal experience and subjective judgment of the physician, and lacks objective diagnostic standards. Therefore, it is urgent to use modern computer technology to conduct quantitative diagnostic research on pulse signals to achieve the standardization and objectivity of pulse diagnosis. Although pulse signal diagnosis technology has achieved certain results, there are still many shortcomings in terms of existing research and application. The main problems include the single pulse signal feature extraction method, the incompleteness of the parameters used in the model, and the imperfect pulse diagnosis algorithm. These factors together lead to a low accuracy rate of pulse diagnosis, so it is urgent to find a more effective solution. Summary of the invention

[0004] In order to overcome the defects of the prior art, the present invention provides a Chinese medicine pulse diagnosis method based on an encoder-decoder framework, which models the pulse signal diagnosis problem as a machine translation problem and establishes a mapping between the pulse signal and the diagnostic text. The present invention aims to solve how to use a deep neural network to extract key features from complex pulse signals, improve the accuracy of pulse signal diagnosis, and directly generate a diagnostic text for the pulse signal.

[0005] A Chinese medicine pulse diagnosis method based on an encoder-decoder framework comprises the following steps:

[0006] Step 1: Collect data: Use different types of sensors to collect pulse signals for people of different physiques and ages, and obtain the diagnosis text of pulse signals by traditional Chinese medicine;

[0007] Step 2: Preprocessing of data set: preprocessing the collected pulse signals to remove background noise, high-frequency noise and baseline drift, and divide the processed pulse signals into training set and test set;

[0008] Step 3: Build a neural network based on the encoder-decoder framework: including the encoder and the decoder. The encoder includes a temporal convolutional network, a long short-term memory neural network, and a fully connected layer. The decoder includes a long short-term memory neural network with an attention mechanism and a fully connected layer.

[0009] The temporal convolutional network receives the input of the encoder to extract multi-scale feature information from the pulse signal, and then sends the convolved sequence to the long short-term memory neural network with several hidden layers for feature processing and enhancement. The encoder output is then calculated through the fully connected layer, and finally the encoder output is parsed through the decoder to obtain the final predicted text.

[0010] Step 4: Training the network: Input the preprocessed training set into the neural network based on the encoder-decoder framework, and generate the predicted text of the health status matching the pulse signal through the encoder and decoder. In the training process of the network model, use the Softmax function as the activation function of the model, and use Adam as the optimization algorithm of the model to optimize the cross entropy loss function. Finally, use Dropout and add an additional auxiliary loss in the training stage to force the model to accurately predict the text representation at the corresponding moment based on the hidden layer representation of the encoder at each time step. After multiple rounds of training, the final network model is obtained;

[0011] Step 5: Testing: Input the preprocessed test set pulse signal into the trained network model to obtain the model's predicted text, compare it with the given real text, and calculate the evaluation index.

[0012] The step 1 specifically includes:

[0013] Different types of sensors are used to collect pulse signals from people with different physiques and different age groups, and the Chinese medicine diagnosis text of the pulse signals is recorded. The pulse signals collected by each tester include the three pulse diagnosis parts of "Cun, Guan, and Chi" with three degrees of depth of "floating, middle, and sinking" on the left and right hands. The specific collection method is to use piezoelectric pressure sensors for young people and adopt a single-point collection method; use photoelectric sensors for children and adopt a single-point collection method; use piezoresistive pressure sensors for the elderly or people with weaker physiques, and adopt multi-point collection and dynamic collection methods.

[0014] The step 2 specifically includes:

[0015] A zero-phase filter was used to remove background noise, and Sym8 wavelet was selected as the wavelet basis function for 7-layer wavelet decomposition. The coefficients of the high-frequency signal after decomposition and the coefficients of the calculated baseline drift signal were set to zero. The signal was then reconstructed to remove high-frequency noise and baseline drift problems, and the data was then divided into training and test sets in a ratio of 8:2.

[0016] The step three specifically includes:

[0017] The temporal convolutional network applies the same temporal short frames at fixed time intervals along the entire input sequence. Setting a step size larger than one sampling point can effectively downsample the output sequence. The network learns a set of such filters to obtain a set of filter sequences that are effectively downsampled to 16Hz and input into the long short-term memory neural network (LSTM).

[0018] The LSTM neural network contains three gating units: forget gate, input gate, and output gate, as well as a cell state. The LSTM neural network uses the output of the temporal convolutional network as the input of the network, processes the input sequence through the gating mechanism, and generates an output sequence containing all time steps.

[0019] The gating mechanism is specifically manifested as follows: first, the sequence data passes through the Sigmoid layer of the forget gate, which is based on the previous time step h t-1 and the current time step x t The input is in the memory cell state C t-1 Each neuron in generates a value between 0 and 1, f t The formula is:

[0020] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0021] The next stage consists of two steps: first, the input gate layer is combined with h t-1 The current input determines the unit state of the current input, and the tanh layer produces Added to the neuron state, the first two processes are combined to provide an updated state value, as follows: t and The formula is:

[0022] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0023]

[0024] Next, update the old neuron state C t-1 To the new neuron state C t , C t-1 Multiply by one f t , then enter Get a measure of how much to update each state value new candidate value, c t The formula is as follows:

[0025]

[0026] Finally, the output of the network is determined. The output is filtered and built on the basis of the neuron state. Specifically, the output component of the neuron is determined by the sigmoid function, and the neuron state is mapped to the range of -1 to 1 using the tanh function. It is multiplied by the above output component. The long short-term memory neural network learns to summarize in a single hidden state to obtain the final o t and h t for:

[0027] o t =σ(W o ·[h t-1 , x t +b o )

[0028] h t =o t *tanh(C t )

[0029] The final encoder hidden state initializes the decoder, which learns to predict the next word in the sequence given the previous word and its own current state. The specific steps are as follows: after the output of the LSTM layer of the decoder, an Attention layer is added. The Attention layer will receive the hidden state of the LSTM neural network layer as input and output a weighted representation. The Attention layer uses a scoring function in the form of a dot product and then uses the softmax function to convert these similarity scores into a probability distribution, namely the Attention weight. The calculated Attention weights are used to weighted sum the output of the LSTM neural network layer to obtain a weighted context vector. The context vector will be used as the input of the fully connected layer of the decoder, and the final output will be generated by the fully connected layer of the decoder.

[0030] The step 4 specifically includes:

[0031] The Softmax function is used as the activation function of the model to map the value of the input sequence to (0, 1). The specific formula is: Adam is used as the optimization algorithm of the model to optimize the cross entropy loss function. The calculation formula of the cross entropy loss is L = -Σy*log(p), where y is the probability distribution of the true label and p is the probability distribution predicted by the model. Dropout randomly sets the output of some neurons to 0 during the entire training phase. The batch size of the temporal convolution is 256, the batch size of the fully connected input layer is 64, and the number of training epochs is 800.

[0032] The entire network is trained simultaneously, extracting features from the text sequence, defining a sliding window on the text sequence, comparing the structure of the text within the window, calculating the ratio of the means, variances, and covariances of the text features within the two windows, and combining brightness, contrast, and structural similarity to calculate the final structural similarity index using the same formula as SSIM:

[0033]

[0034] Among them, μ x and μ y is the mean of the two windows, and is the variance, σ xy is the covariance, c 1 and c 1 It is a constant set to avoid the denominator being zero.

[0035] The step five specifically includes:

[0036] The evaluation indicators include accuracy, recall, and F1 value for evaluating the vocabulary level, and average word error rate for evaluating the sentence level. The accuracy is calculated as the ratio of the number of correctly translated words in the translated text to the total number of words, the recall is calculated as the ratio of the number of correctly translated words in the translated text to the number of words in the reference text, and the F1 value is calculated as the harmonic mean of the accuracy and recall, taking both accuracy and recall into comprehensive consideration.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] The invention relates to a Chinese medicine pulse diagnosis method based on an encoder-decoder framework. The encoder is combined with a temporal convolutional network, an LSTM model and a fully connected layer to enhance feature representation. The decoder adds an Attention mechanism after the LSTM network to significantly improve the focusing ability and performance of the model when processing sequence data. In order to guide the encoder to encode meaningful information, in addition to end-to-end training of the model to decode continuous text from the pulse signal, the article also adds an auxiliary loss in the training stage to force the model to accurately predict the text representation at the corresponding moment based on the hidden layer representation of each time step of the encoder.

[0039] The present invention can effectively capture the key features of pulse signals for predicting health status through deep neural networks, making the diagnosis of pulse signals more accurate.

[0040] The present invention models the pulse signal diagnosis problem as a machine translation problem rather than a simple problem of classifying health status. The model regards the pulse signal as the source language and the corresponding continuous text as the target language, thereby migrating the model method in the field of machine translation to the task of pulse activity decoding, providing new ideas for pulse diagnosis methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0042] Figure 1 A schematic diagram of an encoder-decoder framework in a Chinese medicine pulse diagnosis method based on an encoder-decoder framework of the present invention;

[0043] Figure 2 The internal structure diagram of the long short-term memory neural network in the Chinese medicine pulse diagnosis method based on the encoder-decoder framework of the present invention. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] A Chinese medicine pulse diagnosis method based on an encoder-decoder framework comprises the following steps:

[0046] Step 1: Collect data: Use different types of sensors to collect pulse signals for people of different physiques and ages, and obtain the diagnosis text of pulse signals by traditional Chinese medicine;

[0047] Specifically, different types of sensors were used to collect pulse signals from 500 people of different physiques and ages, and the diagnosis text of Chinese medicine for pulse signals was recorded. The pulse signals collected by each tester included the three pulse diagnosis sites of "Cun, Guan, and Chi" with three depths of "floating, middle, and sinking" on the left and right hands. The specific collection method was to use piezoelectric pressure sensors for young people and a single-point collection method; to use photoelectric sensors for children and a single-point collection method; to use piezoresistive pressure sensors for the elderly or people with weak physiques, and to use multi-point collection and dynamic collection methods.

[0048] Step 2: Preprocessing of data set: preprocessing the collected pulse signals to remove background noise, high-frequency noise and baseline drift, and divide the processed pulse signals into training set and test set;

[0049] The collected matrix pulse signal is preprocessed to remove background noise, high-frequency noise and baseline drift, and the processed pulse signal data set is divided into a training set and a test set.

[0050] Specifically, a zero-phase filter is used to remove background noise, and Sym8 wavelet is selected as the wavelet basis function for 7-layer wavelet decomposition. The coefficients of the high-frequency signal after decomposition and the coefficients of the calculated baseline drift signal are set to zero. The signal is then reconstructed to remove high-frequency noise and baseline drift problems, and the data is then divided into training and test sets in a ratio of 8:2.

[0051] Step 3: Build a neural network based on the encoder-decoder framework: including the encoder and the decoder. The encoder includes a temporal convolutional network, a long short-term memory neural network, and a fully connected layer. The decoder includes a long short-term memory neural network with an attention mechanism and a fully connected layer.

[0052] The temporal convolutional network receives the input of the encoder to extract multi-scale feature information from the pulse signal, and then sends the convolved sequence to the long short-term memory neural network with several hidden layers for feature processing and enhancement. The encoder output is then calculated through the fully connected layer, and finally the encoder output is parsed through the decoder to obtain the final predicted text.

[0053] Specifically, the temporal convolutional network (TCN) applies the same temporal short frame at a fixed time interval (step length) along the entire input sequence, and setting a step length greater than one sampling point can effectively downsample the output sequence. The network learns a set of such filters to obtain a set of filter sequences that are effectively downsampled to 16Hz and input into a long short-term memory neural network (LSTM) to decode continuous text.

[0054] The LSTM unit includes three gating units, namely a forget gate, an input gate and an output gate, and a cell state. The LSTM uses the output of the TCN as the input of the network, processes the input sequence through the gating mechanism, and generates an output sequence containing all time steps.

[0055] The gating mechanism is as follows: first, the sequence data passes through the Sigmoid layer of the forget gate, which is based on the previous time step h. t-1 and the current time step x t The input is in the memory cell state C t-1 Each neuron in generates a value between 0 and 1, f t The formula is:

[0056] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0057] The next stage consists of two steps. First, the input gate layer combines h t-1 And the current input determines the unit state of the current input. Then, the tanh layer produces Added to the neuron state. Afterwards, the first two processes are combined to provide an updated state value. t and The formula is:

[0058] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0059]

[0060] Next, update the old neuron state C t-1 To the new neuron state C t . t-1 Multiply by one f t , then enter Get a measure of how much to update each state value to a new candidate value. t The formula is as follows:

[0061]

[0062] Finally, the output of the network is determined. This output is filtered and built on the basis of the neuron state. Specifically, the output component of the neuron is determined by the sigmoid function, and the neuron state is mapped to the range of -1 to 1 using the tanh function, which is multiplied by the above output component. LSTM learns to summarize them in a single hidden state. The final o t and h t for:

[0063] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0064] h t =o t *tanh(C t )

[0065] The final encoder hidden state initializes the decoder, which learns to predict the next word in the sequence given the previous word and its own current state. The decoder uses a long short-term memory (LSTM-Attention) neural network with an attention mechanism, which is a deep learning model that combines an LSTM network and an attention mechanism. The Attention mechanism is a computational model that simulates the human attention mechanism by calculating the weights of each input sequence element and applying these weights to the input sequence. The specific steps are to add an Attention layer after the output of the LSTM layer of the decoder. This Attention layer will receive the hidden state of the LSTM layer as input and output a weighted representation. The Attention layer usually passes a scoring function in the form of a dot product, and then uses a softmax function to convert these similarity scores into a probability distribution, namely the Attention weight. The output of the LSTM layer is weighted and summed using the calculated Attention weights to obtain a weighted context vector. This context vector will be used as the input of the decoder's fully connected layer. Finally, the fully connected layer of the decoder generates the final output.

[0066] Step 4: Training the network: Input the preprocessed training set into the neural network based on the encoder-decoder framework, and generate the predicted text of the health status matching the pulse signal through the encoder and decoder. In the training process of the network model, use the Softmax function as the activation function of the model, and use Adam as the optimization algorithm of the model to optimize the cross entropy loss function. Finally, use Dropout and add an additional auxiliary loss in the training stage to force the model to accurately predict the text representation at the corresponding moment based on the hidden layer representation of the encoder at each time step. After multiple rounds of training, the final network model is obtained;

[0067] Specifically, the preprocessed training set is input into a neural network based on an encoder-decoder framework, and the predicted text of the health status matching the pulse signal is generated through the encoder and decoder. During the network training process, the Softmax function is used as the activation function of the model to map the value of the input sequence to (0, 1), and the specific formula is: Adam is used as the optimization algorithm of the model to optimize the cross entropy loss function, and the calculation formula of the cross entropy loss is: where y is the probability distribution of the true label and p is the probability distribution predicted by the model; Dropout randomly sets the output of some neurons to 0 during the entire training phase. The batch size of the temporal convolution is 256, the batch size of the fully connected input layer is 64, and the number of training times is 800 epochs.

[0068] An additional auxiliary loss is added during the training phase to force the model to accurately predict the text representation at the corresponding time based on the hidden layer representation of the encoder at each time step. After multiple rounds of training, the final network model is obtained.

[0069] The entire network is trained simultaneously, using the structural similarity index (SSIM) to provide an "auxiliary loss" to guide the network to find a good solution to the word sequence decoding problem. Features (such as word embedding, syntax tree structure, etc.) are extracted from the text sequence, a sliding window is defined on the text sequence, the structure of the text in the window is compared, the ratio of the mean, variance and covariance of the text features in the two windows is calculated, and the brightness, contrast and structural similarity are combined. The final structural similarity index is calculated using the same formula as SSIM:

[0070]

[0071] Among them, μ x and μ y is the mean of the two windows, and is the variance, σ xy is the covariance, c 1 and c 1is a constant set to avoid zero denominator. During testing, SSIM predictions are simply dropped and decoding is based entirely on the decoder output. All training is performed by stochastic gradient descent with backpropagation, with dropout applied to all layers.

[0072] Step 5: Testing: Input the preprocessed test set pulse signal into the trained network model to obtain the model's predicted text, compare it with the given real text, and calculate the evaluation index.

[0073] Specifically, the preprocessed test set pulse signal is input into the trained network model to obtain the model's predicted text, which is compared with the given real text to calculate the evaluation index.

[0074] The evaluation indicators include the accuracy, recall, F1 value for evaluating the vocabulary level and the average (among all sentences) word error rate (WER) for evaluating the sentence level. The accuracy is calculated as the ratio of the number of correctly translated words in the translated text to the total number of words, the recall is calculated as the ratio of the number of correctly translated words in the translated text to the number of words in the reference text, and the harmonic mean of the accuracy and recall is calculated as the F1 value, taking the accuracy and recall into consideration. WER is the minimum number of deletions, insertions, and substitutions required to transform the predicted sentence into a true sentence, normalized by the length of the latter. Therefore, the WER for perfect decoding is 0%, while for incorrect decoding, it is technically unbounded, although in practice it can be limited to 100% by simply predicting empty sentences. For reference, in speech transcription, a WER of 5% is professional level, while 20%-25% is the outer limit of acceptable performance.

[0075] Network training settings:

[0076] The present invention proposes a deep neural network based on an encoder-decoder framework, which is implemented by using PyTorch on a PC equipped with an NVIDIA RTX 4090 GPU. In the training stage, the pulse signal is a feature sequence with the same length, and a variety of data enhancement strategies are adopted, including background denoising, high-frequency denoising, and a method for solving baseline offset. The encoder of the model is initialized by a temporal convolutional network, and feature extraction is performed through an LSTM network, and an attention mechanism is introduced to enhance the model's learning ability for key features. Finally, the output of the encoder is decoded. In the training process of the network model, a Softmax function is used as the activation function of the model, and Adam is used as the optimization algorithm of the model to optimize the cross entropy loss function. Finally, Dropout is used to avoid overfitting problems, wherein the batch size of the temporal convolution is 256, the batch size of the fully connected input layer is 64, and the number of training times is 800 epochs.

[0077] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.

Claims

1. A Chinese medicine pulse diagnosis method based on an encoder-decoder framework, characterized in that: The following steps are involved: Step 1: Collect data: Use different types of sensors to collect pulse signals for people of different physiques and ages, and obtain the diagnosis text of pulse signals by traditional Chinese medicine; Step 2: Preprocessing of data set: preprocessing the collected pulse signals to remove background noise, high-frequency noise and baseline drift, and dividing the processed pulse signals into training set and test set; Step 3: Build a neural network based on the encoder-decoder framework: including the encoder and the decoder. The encoder includes a temporal convolutional network, a long short-term memory neural network, and a fully connected layer. The decoder includes a long short-term memory neural network with an attention mechanism and a fully connected layer. The temporal convolutional network receives the input of the encoder to extract multi-scale feature information from the pulse signal, and sends the convolved sequence to a long short-term memory neural network with several hidden layers for feature processing and enhancement. The output of the encoder is calculated through the fully connected layer, and the output of the encoder is parsed through the decoder to obtain the final predicted text. Step 4: Training the network: Input the preprocessed training set into the neural network based on the encoder-decoder framework, and generate the predicted text of the health status matching the pulse signal through the encoder and decoder. In the training process of the network model, use the Softmax function as the activation function of the model, and use Adam as the optimization algorithm of the model to optimize the cross entropy loss function. Use Dropout to add an additional auxiliary loss in the training stage to force the model to accurately predict the text representation at the corresponding moment based on the hidden layer representation of the encoder at each time step. After multiple rounds of training, the final network model is obtained; Step 5: Test: Input the preprocessed test set pulse signal into the trained network model to obtain the model's predicted text, compare it with the given real text, and calculate the evaluation index.

2. A Chinese medicine pulse diagnosis method based on an encoder-decoder framework according to claim 1, characterized in that: The step 1 specifically includes: Different types of sensors are used to collect pulse signals from people with different physiques and different age groups, and the Chinese medicine diagnosis text of the pulse signals is recorded. The pulse signals collected by each tester include the three pulse diagnosis parts of "Cun, Guan, and Chi" with three degrees of depth of "floating, middle, and sinking" on the left and right hands. The specific collection method is to use piezoelectric pressure sensors for young people and adopt a single-point collection method; use photoelectric sensors for children and adopt a single-point collection method; use piezoresistive pressure sensors for the elderly or people with weak physiques, and adopt multi-point collection and dynamic collection methods.

3. A Chinese medicine pulse diagnosis method based on an encoder-decoder framework according to claim 1, characterized in that: The step 2 specifically includes: A zero-phase filter was used to remove background noise, and Sym8 wavelet was selected as the wavelet basis function for 7-layer wavelet decomposition. The coefficients of the high-frequency signal after decomposition and the coefficients of the calculated baseline drift signal were set to zero. The signal was reconstructed to remove high-frequency noise and baseline drift problems, and then the data was divided into training set and test set in a ratio of 8:

2.

4. A Chinese medicine pulse diagnosis method based on an encoder-decoder framework according to claim 1, characterized in that: The step three specifically includes: The temporal convolutional network applies the same temporal short frames at fixed time intervals along the entire input sequence. Setting a step size larger than one sampling point can effectively downsample the output sequence. The network learns a set of such filters and obtains a set of filter sequences that are effectively downsampled to 16Hz and input into the long short-term memory neural network. The LSTM neural network contains three gating units: forget gate, input gate, and output gate, as well as a cell state. The LSTM neural network uses the output of the temporal convolutional network as the input of the network, processes the input sequence through the gating mechanism, and generates an output sequence containing all time steps. The gating mechanism is specifically manifested as follows: first, the sequence data passes through the Sigmoid layer of the forget gate, which is based on the previous time step h t-1 and the current time step x t The input is in the memory cell state C t-1 Each neuron in generates a value between 0 and 1, f t The formula is: f t =σ(W f ·[h t-1 ,x t ]+b f ) The next stage consists of two steps: first, the input gate layer is combined with h t-1 The current input determines the unit state of the current input, and the tanh layer produces Added to the neuron state, the first two processes are combined to provide an updated state value, as follows: t and The formula is: i t =σ(W i ·[h t-1 ,x t ]+b i ) Next, update the old neuron state C t-1 To the new neuron state C t , C t-1 Multiply by one f t , then enter Get a measure of how much to update each state value new candidate value, c t The formula is as follows: Finally, the output of the network is determined. The output is filtered and built on the basis of the neuron state. Specifically, the output component of the neuron is determined by the sigmoid function, and the neuron state is mapped to the range of -1 to 1 using the tanh function. It is multiplied by the above output component. The long short-term memory neural network learns to summarize in a single hidden state to obtain the final o t and h t for: the t =σ(W o ·[h t-1 ,x t ]+b o ) h t =o t *tanh(C t ) The final encoder hidden state initializes the decoder, which learns to predict the next word in the sequence given the previous word and its own current state. The specific steps are as follows: after the output of the LSTM layer of the decoder, an Attention layer is added. The Attention layer will receive the hidden state of the LSTM neural network layer as input and output a weighted representation. The Attention layer uses a scoring function in the form of a dot product and then uses the softmax function to convert these similarity scores into a probability distribution, namely the Attention weight. The calculated Attention weights are used to weighted sum the output of the LSTM neural network layer to obtain a weighted context vector. The context vector will be used as the input of the fully connected layer of the decoder, and the final output will be generated by the fully connected layer of the decoder.

5. a kind of Chinese medicine pulse diagnosis method based on encoder-decoder framework according to claim 1, it is characterized in that: The step 4 specifically includes: The Softmax function is used as the activation function of the model to map the value of the input sequence to (0, 1). The specific formula is: Adam is used as the optimization algorithm of the model to optimize the cross entropy loss function. The calculation formula of the cross entropy loss is L = -Σy*log(p), where y is the probability distribution of the true label and p is the probability distribution predicted by the model. Dropout randomly sets the output of some neurons to 0 during the entire training phase. The batch size of the temporal convolution is 256, the batch size of the fully connected input layer is 64, and the number of training epochs is 800. The entire network is trained simultaneously, extracting features from the text sequence, defining a sliding window on the text sequence, comparing the structure of the text within the window, calculating the ratio of the means, variances, and covariances of the text features within the two windows, and combining brightness, contrast, and structural similarity to calculate the final structural similarity index using the same formula as SSIM: Among them, μ x and μ y is the mean of the two windows, and is the variance, σ xy is the covariance, c1 and c2 are constants set to avoid the denominator being zero.

6. A Chinese medicine pulse diagnosis method based on an encoder-decoder framework according to claim 1, characterized in that: The step five specifically includes: The evaluation indicators include accuracy, recall, and F1 value for evaluating the vocabulary level, and average word error rate for evaluating the sentence level. The accuracy is calculated as the ratio of the number of correctly translated words in the translated text to the total number of words, the recall is calculated as the ratio of the number of correctly translated words in the translated text to the number of words in the reference text, and the F1 value is calculated as the harmonic mean of the accuracy and recall, taking both accuracy and recall into comprehensive consideration.

Images