Intelligent scheme recommendation system and method based on myoelectric pressure and deep learning

Abstract

The invention discloses an intelligent scheme recommendation system and method based on electromyographic pressure and deep learning, and relates to the technical field of deep learning. The intelligent scheme recommendation system synchronously collects pelvic floor electromyographic signals and pressure signals through an electromyographic pressure integrated electrode, extracts evaluation data of all stages such as a pre-resting stage and a fast muscle stage, and sends the evaluation data to a server; the method comprises the following steps: constructing a feature vector by combining personal information of a user, after standardization, mapping the feature into a high-dimensional feature by using an encoder, inputting the high-dimensional feature into a decoder through sine and cosine position encoding, and generating a scheme recommendation matrix by the decoder; in addition, in model training, the model is optimized through a cross entropy loss function; through multi-data fusion and intelligent scheme generation, the pelvic floor training efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning, and specifically to a scheme intelligent recommendation system and method based on electromyogram pressure and deep learning. Background Art

[0002] The pelvic floor muscles are a group of muscles located at the bottom of the pelvis, shaped like a "hanging net", mainly responsible for supporting pelvic organs, controlling excretion functions, and participating in core stability. The health of the pelvic floor muscles directly affects the quality of life. Once the elasticity of this "net" deteriorates and the "hanging force" is insufficient, the organs within the "net" cannot be maintained in their normal positions. Therefore, it is necessary to exercise the pelvic floor muscles to solve pelvic floor muscle dysfunction.

[0003] The existing pelvic floor muscle training programs are mainly formulated by doctors according to the user's pelvic floor muscle conditions. This way of formulating programs based on pelvic floor muscle conditions has the following problems: First, data resources are not fully utilized, such as the user's personal information data and pelvic floor muscle assessment report data; second, after completing the training according to the formulated program, the conditions of the pelvic floor muscles are not continuously tracked and fed back, and the formulated program cannot be evaluated; third, the levels of doctors vary, and judgments are made based on their own work experience, which may lead to the formulated program not fully meeting the user's own needs, missing the best training time, and thus resulting in low efficiency of pelvic floor muscle training. Summary of the Invention

[0004] The purpose of the present invention is to provide a scheme intelligent recommendation system and method based on electromyogram pressure and deep learning to solve the problems raised in the prior art.

[0005] To achieve the above purpose, the present invention provides the following technical solution: A scheme intelligent recommendation system based on electromyogram pressure and deep learning, the recommendation system includes: a data acquisition module, a data processing module, a position encoding module, an encoder module, and a decoder module; The data acquisition module is used to collect the pelvic floor electromyogram signal and pressure signal of the user, extract the evaluation data, and collect the user's personal information; The data processing module is used to perform standardization processing on the evaluation data and personal information, construct a feature vector, and unify the preset component information into a matrix format; The position encoding module is used to encode the position of each element in the target matrix to generate a position encoding matrix; The encoder module is used to map the feature vector into a high-dimensional feature vector and expand it into a feature matrix; The decoder module is used to fuse the features in the feature matrix with the position encoding to generate a scheme recommendation matrix.

[0006] Furthermore, the data acquisition module includes: Synchronously collect pelvic floor electromyogram signals and pressure signals through an integrated electromyogram-pressure electrode; extract evaluation data for each stage from the collected signals using the mathematical formulas of mean and standard deviation. The evaluation data for each stage includes: The evaluation data of static pelvic floor muscle tension in the pre-rest stage includes pre-rest mean value, variability, and comprehensive score; The evaluation data of dynamic fast muscle strength and reaction speed in the fast muscle stage includes fast muscle maximum value, fast muscle contraction time, fast muscle relaxation time, fast muscle electromyogram score, and fast muscle pressure score; The evaluation data of dynamic slow muscle strength and stability of muscle endurance contraction control in the slow muscle stage includes slow muscle electromyogram mean value, slow muscle pressure mean value, slow muscle variability, slow muscle rise time, slow muscle recovery time, slow muscle electromyogram score, and slow muscle pressure score; The evaluation data of stability of muscle endurance contraction control in the endurance stage includes endurance electromyogram maximum value, endurance pressure maximum value, endurance rise time, endurance recovery time, endurance electromyogram score, and endurance pressure score; The evaluation data of static pelvic floor muscle tension in the post-rest stage includes post-rest mean value, variability, and comprehensive score; The personal information includes BMI, age at delivery, date of delivery, mode of delivery, number of pregnancies, and baby weight.

[0007] Furthermore, the data processing module includes: Taking the evaluation data and personal information as input feature values, perform standardization processing on the input feature values using the Z-score standardization formula. The Z-score standardization formula is: ; where x represents the input feature value, x norm represents the input feature matrix after standardization processing, represents the standard deviation, represents the mean value; Arrange the feature values corresponding to the evaluation data in sequence according to the stage order to obtain the first target vector, arrange the feature values corresponding to the personal information in sequence to obtain the second target vector, and splice the first target vector and the second target vector to obtain a feature vector; In the above steps, since the units of evaluation data (such as electromyogram values, pressure values, scores, etc.) and personal information (BMI, age at delivery, etc.) are completely different and the numerical differences are large, it is necessary to eliminate the influence of dimensions to ensure the stability of the numerical values and avoid serious model fitting deviation caused by extreme values; the above steps also preprocess the data, discard a small amount of missing data in the evaluation report, and avoid the inaccurate individual evaluation data affecting the overall effect of the model.

[0008] Taking each preset component as an element, arranging the elements according to the preset number of rows and columns, and padding with 0 when the number of elements in each row is not equal to the number of columns, to obtain a target matrix that meets the preset number of rows and columns; In the above steps, to improve the model training efficiency, the scheme information is all composed of components. Different components have different training effects on the pelvic floor muscles, such as type I stimulation 1, type I stimulation 2, muscle contraction and relaxation training 1, muscle contraction and relaxation training 2, etc.

[0009] Further, the position encoding module includes: Performing position encoding on the positions of the elements in the even columns of the target matrix using the sine function, with the formula as follows: ; where PE (pos,2i) represents the element in the pos-th row and 2i-th column of the position encoding matrix, and d model is the hidden layer dimension of the Transformer model; Performing position encoding on the positions of the elements in the odd columns of the target matrix using the cosine function, with the formula as follows: ; where PE (pos,2i+1) represents the position encoding of the element in the pos-th row and (2i + 1)-th column of the position encoding matrix; In the above steps, when generating the scheme recommendation matrix, the model needs to understand the order and hierarchical relationship of the components in different positions in the scheme (such as the semantic differences between the 1st row and the 15th row, and the 3rd column and the 5th column). In conventional sequence processing models, the original attention mechanism cannot perceive the relative position information of the elements in the sequence, so the position information of the elements is added here; the present invention adopts a position encoding scheme combining sine and cosine functions. By integrating the row and column position information into a single encoding, the model can distinguish the semantic differences between different rows and columns in the matrix.

[0010] Further, the encoder module includes: Inputting the feature vector into a linear transformation layer for linear transformation, in the following manner: ; where H enc represents the feature vector after linear transformation, W e represents the weight matrix of linear transformation, and b e represents the bias vector of linear transformation; Processing the feature after linear transformation through a non-linear activation function to obtain an extended feature, denoted as ; where, represents the extended feature, Denote the eigenvector after linear transformation, \(n\) represents the number of rows of the eigenvector, and \(m\) represents the number of columns of the eigenvector; the non-linear transformation method is as follows: ; Among them, when the input feature \(x\geq0\), the output is \(f(x) = x\); when the input feature \(x\lt0\), the output is \(f(x) = 0\); In the above steps, the low-dimensional features are converted into high-dimensional features. The purpose of this dimension elevation is to enable the model to capture the complex relationships between features; the above non-linear activation function will screen and reorganize the features, highlighting important information and suppressing irrelevant information; when a negative number is input, the output is "truncated" to 0, suppressing invalid signals; when a positive number is input, the output is directly equal to the input, retaining valid signals; the features after non-linear transformation will be standardized again to ensure that the distribution of the features is more stable, which helps the stability and convergence speed of model training.

[0011] Furthermore, the decoder module includes: a self-attention layer, a cross-attention layer, a feed-forward network layer, a component prediction layer, and a probability matrix reshaping layer; The self-attention layer generates a corresponding position index for each position in the position encoding matrix, and converts each position index into a vector through the embedding layer, denoted as ; superimpose the position encoding corresponding to each element in the position encoding matrix with the position index to generate ; where, \(PE\) represents the position encoding corresponding to each element, represents the currently generated position; In the above steps, the generated enables the model to perceive the position order; The cross-attention layer fuses the features in the feature matrix with the position encoding through the cross-attention mechanism. The formula is: ; where, \(Q\) represents the input of the current layer of the decoder , \(K\) and \(V\) both represent the extended features of the encoder module , represents the degree of attention of the current position to all positions; In the above steps, the decoder uses the cross-attention mechanism to enable each target position \(p\) to "query" the encoder memory, that is, the generated by the encoder; where, \(Q\) (Query, query) represents "what information is needed at the current position", and \(K\) (Key, key), \(V\) (Value, value) represent "the information stored in the encoder memory"; The feed-forward network layer raises the dimension of the output result of the attention mechanism through non-linear transformation, and reduces the dimension of the output result after dimension elevation through linear transformation to obtain a feature matrix after removing redundant information; The component prediction layer maps the output result of the feed-forward network to the component candidate space through non-linear transformation, and generates a probability distribution through Softmax. The formula is as follows: ; Among them, represents the probability of selecting the k-th component at the j-th position of the i-th user, and N represents the number of components in the candidate space; The probability matrix reshaping layer divides the fused feature matrix into n rows and m columns, transforms it into an n×m matrix, and selects the component with the highest predicted probability in the component list to fill the feature matrix, obtaining the solution recommendation matrix.

[0012] Furthermore, to better implement the above system, a solution intelligent recommendation method based on electromyogram pressure and deep learning is also provided. The recommendation method includes: Step S1: Collect the pelvic floor electromyogram signal and pressure signal of the user through the electromyogram pressure integrated electrode, extract the evaluation data of each stage through the mean value, standard deviation and variability, and collect the personal information of the user; Step S2: Clean the collected data, standardize the cleaned evaluation data and personal information, construct a feature vector, and organize the preset component information into a matrix format to obtain the target matrix; Step S3: Perform sine coding on the even-column elements and cosine coding on the odd-column elements in the target matrix to obtain the position coding matrix; Step S4: Perform linear transformation and non-linear transformation on the feature vector in sequence to obtain the extended feature vector; Step S5: Fuse the extended feature vector and the position coding matrix through the decoder to obtain the solution recommendation matrix.

[0013] Furthermore, step S1 includes: The calculation formula for the mean value is: ; Among them, represents the mean value, and x i represents the i-th sample value collected, and N represents the number of samples collected in each stage; The calculation formula for the standard deviation is: ; Among them, represents the standard deviation; The calculation formula for the variability is: ; Among them, CV represents the variability.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By synchronously collecting multi-dimensional signals through an integrated electromyogram-pressure electrode, the comprehensiveness and accuracy of pelvic floor assessment are improved.

[0015] 2. By combining the user's personal information with a deep learning algorithm, a personalized pelvic floor muscle training plan is generated, improving the training efficiency and effect.

[0016] 3. By adopting positional encoding and an attention mechanism, the rationality of the structure and order of the recommended plan is ensured. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of the system structure of a system and method for intelligent recommendation of a plan based on electromyogram pressure and deep learning according to the present invention; Figure 2 It is a schematic diagram of the method flow of a system and method for intelligent recommendation of a plan based on electromyogram pressure and deep learning according to the present invention; Figure 3 It is an evaluation diagram of a pelvic floor electromyogram report of a system and method for intelligent recommendation of a plan based on electromyogram pressure and deep learning according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0019] Embodiment 1: As Figure 1 shown, the present invention provides a technical solution, an intelligent recommendation system for a plan based on electromyogram pressure and deep learning, the system includes: a data acquisition module, a data processing module, a positional encoding module, an encoder module, and a decoder module; The data acquisition module is used to collect the pelvic floor electromyogram signal and pressure signal of the user, extract evaluation data and collect the personal information of the user; The data processing module is used to perform standardization processing on the evaluation data and personal information, construct a feature vector and unify the preset component information into a matrix format; The positional encoding module is used to encode the position of each element in the target matrix to generate a positional encoding matrix; The encoder module is used to map the feature vector into a high-dimensional feature vector and expand it into a feature matrix; The decoder module is used to fuse the features in the feature matrix with the positional encoding to generate a plan recommendation matrix; Among them, the data acquisition module includes: Synchronously collect pelvic floor electromyogram signals and pressure signals through an integrated electromyogram and pressure electrode; extract the evaluation data of each stage from the collected signals using the mathematical formulas of mean value and standard deviation; In an embodiment of the present invention, pelvic floor evaluation is performed through an integrated electromyogram and pressure pelvic floor electrode. The sampling rate is 128 times per second, for a total of 6 minutes and 17 seconds. A total of 128×377 = 48,256 electromyogram signals are collected, and 48,256 pressure signals are synchronously collected; Refer to Figure 3 , the evaluation data of each stage includes: The evaluation data of static pelvic floor muscle tension in the pre-rest stage includes pre-rest mean value, variability, and comprehensive score; The evaluation data of dynamic fast muscle strength and reaction speed in the fast muscle stage includes fast muscle maximum value, fast muscle contraction time, fast muscle relaxation time, fast muscle electromyogram score, and fast muscle pressure score; The evaluation data of dynamic slow muscle strength and stability of muscle endurance contraction control in the slow muscle stage includes slow muscle electromyogram mean value, slow muscle pressure mean value, slow muscle variability, slow muscle rise time, slow muscle recovery time, slow muscle electromyogram score, and slow muscle pressure score; The evaluation data of stability of muscle endurance contraction control in the endurance stage includes endurance electromyogram maximum value, endurance pressure maximum value, endurance rise time, endurance recovery time, endurance electromyogram score, and endurance pressure score; The evaluation data of static pelvic floor muscle tension in the post-rest stage includes post-rest mean value, variability, and comprehensive score; Personal information includes BMI, age at delivery, date of delivery, mode of delivery, number of pregnancies, and baby weight; In an embodiment of the present invention, the personal information of the user is collected in the form of a form for comprehensive evaluation of the pelvic floor muscle condition, including BMI, age at delivery, date of delivery, mode of delivery, number of pregnancies, and baby weight, a total of 6 items; Among them, the data processing module includes: Taking the evaluation data and personal information as input feature values, perform standardization processing on the input feature values using the Z-score standardization formula. The Z-score standardization formula is: ; Among them, x represents the input feature value, x norm represents the input feature matrix after standardization processing, represents the standard deviation, represents the mean value; Arrange the feature values corresponding to the evaluation data in sequence according to the stage order to obtain the first target vector, arrange the feature values corresponding to the personal information in sequence to obtain the second target vector, and splice the first target vector and the second target vector to obtain the feature vector; Taking each preset component as an element, arranging the elements according to the preset number of rows and columns, and padding with 0 when the number of elements in each row is not equal to the number of columns to obtain a target matrix that meets the preset number of rows and columns; In the embodiments of the present invention, to ensure the effectiveness of the solution, a total of 300 components are formulated. Each solution has 15 sections, and each section consists of 1 to 5 components, that is, matrix information of 15×5. To ensure the optimal treatment effect, the training data collected in the present invention are all excellent training solutions ranked in the top 20% in terms of training effect. The training effect evaluation method is: the top 20% with the highest increase in the comprehensive score of the pelvic floor evaluation report before and after the completion of the training solution. To unify the input content of the model, when there are less than 5 components, padding with 0 is performed to ensure that the output is a 15×5 matrix; Among them, the position encoding module includes: Performing position encoding on the positions of the elements in the even columns of the target matrix using the sine function, and the formula is as follows: ; Among them, PE (pos,2i) represents the element in the pos-th row and the 2i-th column of the position encoding matrix, and d model is the hidden layer dimension of the Transformer model, which is designed as 128 dimensions in the present invention; Performing position encoding on the positions of the elements in the odd columns of the target matrix using the cosine function, and the formula is as follows: ; Among them, PE (pos,2i+1) represents the position encoding of the element in the pos-th row and the 2i + 1-th column of the position encoding matrix; Among them, the encoder module includes: Inputting the feature vector into a linear transformation layer for linear transformation, and the method is as follows: ; Among them, H enc represents the feature vector after linear transformation, W e represents the weight matrix of linear transformation, and b e represents the bias vector of linear transformation; Processing the feature after linear transformation through a non-linear activation function to obtain an extended feature, denoted as ; Among them, represents the extended feature, represents the feature vector after linear transformation, n represents the number of rows of the feature vector, and m represents the number of columns of the feature vector; the non-linear transformation method is as follows: ; Among them, when the input feature x ≥ 0, the output is f(x) = x; when the input feature x < 0, the output is f(x) = 0; Among them, the decoder module includes: a self-attention layer, a cross-attention layer, a feed-forward network layer, a component prediction layer, and a probability matrix reshaping layer; The self-attention layer generates a corresponding position index for each position in the position encoding matrix, and converts each position index into a vector through an embedding layer, denoted as ; The position encoding corresponding to each element in the position encoding matrix is superimposed with the position index to generate ; Among them, PE represents the position encoding corresponding to each element, represents the currently generated position; The cross-attention layer fuses the features in the feature matrix with the position encoding through a cross-attention mechanism, and the formula is: ; Among them, Q represents the input of the current layer of the decoder , K and V both represent the features after expansion of the encoder module , represents the degree of attention of the current position to all positions; The feed-forward network layer raises the dimension of the output result of the attention mechanism through a non-linear transformation, and reduces the dimension of the output result after dimension raising through a linear transformation to obtain a feature matrix after removing redundant information; The component prediction layer maps the feature dimension to the component candidate space through a non-linear transformation of the output result of the feed-forward network, and generates a probability distribution through Softmax. The formula is as follows: ; Among them, represents the probability of selecting the k-th component at the j-th position of the i-th user; In the embodiment of the present invention, during the training phase of the model, a cross-entropy loss function will be added, and its formula is: ; Among them, represents the loss function value, which measures the difference between the model prediction and the true value, and optimizes the model parameters by minimizing ; N represents the batch size, that is, the number of samples included in a training batch; 75 represents the sequence length of each sample, corresponding to 75 positions after the 15×5 matrix is unfolded; represents the double summation operation, with the outer layer traversing each sample n in the batch and the inner layer traversing the 75 positions p of each sample; represents the true component label at the p-th position of the n-th sample (the component id with a value of (1~300) or 0 represents padding); X(n) represents the input features of the n-th sample (i.e., a 30-dimensional user feature vector); represents that when the model takes the input X (n) , the probability that the p-th position of the n-th sample is the true label ; represents the natural logarithm of the above probability, which converts the probability to a logarithmic scale for easy calculation and optimization; represents taking the average of the logarithmic probabilities over all samples and positions and adding a negative sign. The negative sign makes the loss function become "negative log-likelihood", and the model predicts the probability of the true label by minimizing the loss; The above cross-entropy loss function calculates the loss by taking the average negative value of the logarithmic probability of the model predicting the true label, driving the model to adjust parameters to increase the probability of predicting the true label, thereby optimizing the model performance; The probability matrix reshaping layer divides the fused feature matrix into n rows and m columns, transforms it into an n×m matrix, and in the component list, selects the component with the highest predicted probability to fill the feature matrix to obtain the solution recommendation matrix; Example 2: As Figure 2 shown, in order to better implement the above system, a solution intelligent recommendation method based on electromyogram pressure and deep learning is also provided. The recommendation method includes: Step S1: Collect the pelvic floor electromyogram signal and pressure signal of the user through an electromyogram pressure integrated electrode, extract the evaluation data of each stage through the mean value, standard deviation and variability, and collect the personal information of the user; Step S2: Clean the collected data, standardize the cleaned evaluation data and personal information, construct a feature vector, and organize the preset component information into a matrix format to obtain a target matrix; Step S3: Perform sine encoding on the elements in the even columns of the target matrix and cosine encoding on the elements in the odd columns to obtain a position encoding matrix; Step S4: Perform a linear transformation and a non-linear transformation on the feature vector in sequence to obtain an extended feature vector; Step S5: Fuse the extended feature vector and the position encoding matrix through a decoder to obtain a solution recommendation matrix; Among them, Step S1 includes: The calculation formula for the mean value is: ; Among them, represents the mean value, x i represents the i-th sample value collected, and N represents the number of samples collected in each stage; The calculation formula for the standard deviation is: ; Among them, represents the standard deviation; The calculation formula of variability is: ; where CV represents variability.

[0020] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or equivalently replace some of the technical features. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A scheme intelligent recommendation system based on electromyogram pressure and deep learning, characterized in that: The recommendation system includes: a data acquisition module, a data processing module, a position encoding module, an encoder module, and a decoder module; The data acquisition module is used to collect the pelvic floor electromyogram signal and pressure signal of the user, extract the evaluation data, and collect the personal information of the user; The data processing module is used to perform standardization processing on the evaluation data and personal information, construct a feature vector, and unify the preset component information into a matrix format; The position encoding module is used to encode the position of each element in the target matrix to generate a position encoding matrix; The encoder module is used to map the feature vector into a high-dimensional feature vector and expand it into a feature matrix; The decoder module is used to fuse the features in the feature matrix with the position encoding to generate a solution recommendation matrix.

2. The intelligent recommendation system for solutions based on electromyogram pressure and deep learning according to claim 1, characterized in that: The data acquisition module includes: Synchronously collect the pelvic floor electromyogram signal and pressure signal through an integrated electromyogram and pressure electrode; use the mathematical formulas of the average value and standard deviation to extract the evaluation data of each stage from the collected signals; The evaluation data of each stage includes: The evaluation data of the static pelvic floor muscle tension in the pre-rest stage includes the pre-rest average value, variability, and comprehensive score; The evaluation data of the dynamic fast muscle strength and reaction speed in the fast muscle stage includes the fast muscle maximum value, fast muscle contraction time, fast muscle relaxation time, fast muscle electromyogram score, and fast muscle pressure score; The evaluation data of the dynamic slow muscle strength and the stability of muscle endurance contraction control in the slow muscle stage includes the slow muscle electromyogram average value, slow muscle pressure average value, slow muscle variability, slow muscle rise time, slow muscle recovery time, slow muscle electromyogram score, and slow muscle pressure score; The evaluation data of the stability of muscle endurance contraction control in the endurance stage includes the endurance electromyogram maximum value, endurance pressure maximum value, endurance rise time, endurance recovery time, endurance electromyogram score, and endurance pressure score; The evaluation data of the static pelvic floor muscle tension in the post-rest stage includes the post-rest average value, variability, and comprehensive score; The personal information includes BMI, delivery age, delivery date, delivery method, number of pregnancies, and baby weight.

3. The intelligent recommendation system based on electromyogram pressure and deep learning according to claim 1, characterized in that: The data processing module includes: Taking the evaluation data and personal information as input feature values, and performing standardization processing on the input feature values using the Z-score standardization formula. The Z-score standardization formula is: ； Among them, x represents the input eigenvalue, and x norm represents the input feature matrix after normalization processing, represents the standard deviation, represents the average value; Sequentially arranging the feature values corresponding to the evaluation data in the order of stages to obtain a first target vector, sequentially arranging the feature values corresponding to the personal information to obtain a second target vector, and splicing the first target vector and the second target vector to obtain a feature vector; Taking each preset component as an element, arranging the elements according to the preset number of rows and columns, and performing zero-padding when the number of elements in each row is not equal to the number of columns to obtain a target matrix that meets the preset number of rows and columns.

4. The intelligent recommendation system for solutions based on EMG pressure and deep learning according to claim 1, characterized in that: The position encoding module includes: Using the sine function to perform position encoding on the positions of the elements in the even columns of the target matrix. The formula is as follows: ； Among them, PE (pos,2i) represents the element in the \(pos\)-th row and the \(2i\)-th column of the position encoding matrix, and \(d\) model is the hidden layer dimension of the Transformer model; Using the cosine function to perform position encoding on the positions of the elements in the odd columns of the target matrix. The formula is as follows: ； Among them, PE (pos,2i+1) represents the position encoding of the element in the (pos)-th row and the (2i + 1)-th column in the position encoding matrix.

5. The intelligent recommendation system for solutions based on electromyogram pressure and deep learning according to claim 1, characterized in that: The encoder module includes: Inputting the feature vector into a linear transformation layer for linear transformation. The method is as follows: ； Among them, H enc represents the eigenvector after linear transformation, W e represents the weight matrix of linear transformation, b e represents the bias vector of linear transformation; The features after linear transformation are processed by a non-linear activation function to obtain the extended features, denoted as ; where represents the extended features,[[]] represents the feature vector after linear transformation, n represents the number of rows of the feature vector, and m represents the number of columns of the feature vector; the non-linear transformation method is as follows: ； Among them, when the input feature x ≥ 0, the output is f(x) = x; when the input feature x < 0, the output is f(x) = 0.

6. The intelligent recommendation system for solutions based on electromyogram pressure and deep learning according to claim 1, characterized in that: The decoder module includes: a self-attention layer, a cross-attention layer, a feed-forward network layer, a component prediction layer, and a probability matrix reshaping layer; The self-attention layer generates a corresponding position index for each position in the position encoding matrix, and converts each position index into a vector through an embedding layer, denoted as ; superimpose the position encoding corresponding to each element in the position encoding matrix with the position index to generate ; where, PE represents the position encoding corresponding to each element, represents the currently generated position; The cross-attention layer fuses the features in the feature matrix with the position encoding through the cross-attention mechanism. The formula is: ； Among them, Q represents the input of the current layer of the decoder , both K and V represent the features after the expansion of the encoder module , represents the degree of attention of the current position to all positions; The feed-forward network layer raises the dimension of the output result of the attention mechanism through a non-linear transformation, and then reduces the dimension of the raised output result through a linear transformation to obtain a feature matrix after removing redundant information; The component prediction layer maps the feature dimension to the component candidate space through a non-linear transformation of the output result of the feed-forward network, and generates a probability distribution through Softmax. The formula is as follows: ； Among them, represents the probability of selecting the k-th component at the j-th position of the i-th user, and N represents the number of components in the candidate space; The probability matrix reshaping layer divides the fused feature matrix into n rows and m columns, transforms it into an n×m matrix, and selects the component with the highest prediction probability in the component list to fill the feature matrix to obtain a solution recommendation matrix.

7. A method for intelligent recommendation of solutions based on electromyogram pressure and deep learning, which is used to execute an intelligent recommendation system for solutions based on electromyogram pressure and deep learning described in claims 1-6, and is characterized in that: The recommendation method includes: Step S1: Collect the pelvic floor electromyogram signal and pressure signal of the user through an electromyogram-pressure integrated electrode, extract the evaluation data of each stage through the mean value, standard deviation, and variability, and collect the personal information of the user; Step S2: Clean the collected data, standardize the cleaned evaluation data and personal information, construct a feature vector, and organize the preset component information into a matrix format to obtain a target matrix; Step S3: Perform sine encoding on the elements in the even columns of the target matrix and cosine encoding on the elements in the odd columns to obtain a position encoding matrix; Step S4: Perform a linear transformation and then a non-linear transformation on the feature vector in sequence to obtain an extended feature vector; Step S5: Fuse the extended feature vector and the position encoding matrix through a decoder to obtain a solution recommendation matrix.

8. The intelligent recommendation method for solutions based on electromyogram pressure and deep learning according to claim 7, characterized in that: Step S1 includes: The calculation formula for the mean value is: ； Among them, represents the average value, and x i represents the i-th sampled value collected, and N represents the number of samples collected in each stage; The calculation formula for the standard deviation is: ； Among them, represents the standard deviation; The calculation formula for the variability is: ； Among them, CV represents the variability.

Images