Uterine electromyographic signal generation method based on depth generation model

Abstract

The invention discloses a uterus electromyographic signal generation method based on a deep generative model, which comprises the following steps: 1, acquiring uterus electromyographic signals of the abdomen of a pregnant woman through a multi-channel surface electrode, and carrying out filtering, noise reduction and normalization preprocessing; 2, performing short-time Fourier transform on the preprocessed signal to obtain time-frequency representation, and decomposing the time-frequency representation into low-frequency and high-frequency components; 3, respectively establishing reconstruction processes of the low-frequency component and the high-frequency component, and obtaining discrete potential representation through training; 4, on the basis of an encoder and a decoder which are subjected to reconstruction training, modeling is conducted on the low-frequency discrete sequence and the high-frequency discrete sequence through a bidirectional Transform prior model, and potential space distribution of the low-frequency discrete sequence and the high-frequency discrete sequence is learned; and 5, in a generation stage, sampling through a priori model to obtain a token sequence, and generating a high-quality uterine myoelectricity signal after decoding and inverse short-time Fourier transform. The signal generated by the method has high authenticity and stability while maintaining the consistency of the time sequence characteristics and the frequency spectrum, and can be applied to uterine contraction monitoring and premature delivery risk prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of uterine electromyography signal processing and analysis, and particularly relates to a uterine electromyography signal generation method using a deep generative model BACKGROUND

[0002] Electrohysterogram (EHG) is an electrophysiological signal reflecting the electrical activity of the uterine smooth muscle of pregnant women, which can directly reveal the occurrence and intensity of uterine contractions. A large number of clinical studies have shown that EHG has important application value in labor monitoring, uterine contraction detection, and preterm birth risk prediction. However, due to individual differences among pregnant women, limitations of clinical collection conditions, and the fact that the signal is easily disturbed by noise, it is difficult to obtain high-quality and large-scale uterine electromyography signal data. At the same time, the number of EHG samples of different categories is obviously imbalanced, especially the abnormal samples are extremely scarce, which to some extent restricts the algorithm research based on EHG and its popularization and application in the clinic.

[0003] In view of the insufficient EHG sample quantity and the unbalanced category distribution, the existing technology often uses synthetic sampling methods for data augmentation. Such methods mainly rely on the distribution of a small number of original class samples in the feature space, and synthesize new data samples through synthetic sampling algorithms. Among them, the commonly used algorithms include SMOTE (Synthetic Minority Over-sampling Technique), ADASYN (Adaptive Synthetic Sampling) and their improved methods. Although the above methods can alleviate the sample imbalance problem to some extent, the generated samples often have two limitations: first, the newly synthesized samples may weaken the class distinguishing features of the original samples, thereby reducing the discriminant ability of the model; second, some synthetic samples are less different from the real samples, and even approximate replication, which can easily lead to overfitting of the classification model in training.

[0004] With the development of deep neural networks and artificial intelligence generation models, researchers have begun to explore the use of deep generation methods to synthesize time series data. Since uterine electromyography signals belong to one-dimensional time series, their generation task can be regarded as a time series generation problem. In recent years, methods based on deep generative models such as generative adversarial networks (GAN) and recurrent neural networks (RNN) have been applied to time series generation and have made some progress in signal authenticity and diversity. However, this kind of method still has shortcomings: the training process of GAN model is unstable and easy to cause mode collapse, RNN model is difficult to effectively model long-range dependencies, in addition, the existing methods have limited performance in generating time-frequency domain consistency, so it is difficult to be directly applied to high-fidelity generation of uterine electromyography signals. SUMMARY

[0005] The present application is to solve the above-mentioned deficiencies in the prior art, a uterine myoelectric signal generation method based on deep generative model is proposed, in order to improve the authenticity, stability and diversity of the synthesized signal, to solve the current uterine myoelectric signal scarcity and sample imbalance problem, at the same time overcome the long-range time sequence dependence difficult to model and the generation signal time-frequency domain consistency is insufficient and other problems in the prior art in the generation of uterine myoelectric signal, so as to provide reliable data support for labor monitoring, uterine contraction detection and premature birth risk prediction.

[0006] The present application is to solve the above-mentioned deficiencies in the prior art, a uterine myoelectric signal generation method based on deep generative model is proposed, in order to improve the authenticity, stability and diversity of the synthesized signal, to solve the current uterine myoelectric signal scarcity and sample imbalance problem, at the same time overcome the long-range time sequence dependence difficult to model and the generation signal time-frequency domain consistency is insufficient and other problems in the prior art in the generation of uterine myoelectric signal, so as to provide reliable data support for labor monitoring, uterine contraction detection and premature birth risk prediction. The present application is to solve the above-mentioned deficiencies in the prior art, a uterine myoelectric signal generation method based on deep generative model is proposed, in order to improve the authenticity, stability and diversity of the synthesized signal, to solve the current uterine myoelectric signal scarcity and sample imbalance problem, at the same time overcome the long-range time sequence dependence difficult to model and the generation signal time-frequency domain consistency is insufficient and other problems in the prior art in the generation of uterine myoelectric signal, so as to provide reliable data support for labor monitoring, uterine contraction detection and premature birth risk prediction. Step 1: uterine myoelectric signals of pregnant women's abdomen are collected by m-channel surface electrodes and pretreated to obtain a pretreated uterine myoelectric signal sequence , let Any channel of the pretreated uterine myoelectric signal is denoted as ; m represents the number of channels; Step 2: short-time Fourier transform is performed on to obtain a time-frequency representation u, and the u is band-decomposed according to frequency components to obtain a low-frequency component and a high-frequency component ; after inverse short-time Fourier transform is performed on and , low-frequency time-domain signals and high-frequency time-domain signals are obtained accordingly; Step 3: a vectorization reconstruction network is constructed, including: a low-frequency encoder and a high-frequency encoder , a discrete coding unit (a low-frequency codebook and a high-frequency codebook ), a low-frequency decoder and a high-frequency decoder , and and are encoded, quantized and decoded to obtain reconstructed low-frequency time-domain signals and reconstructed high-frequency time-domain signals , and to form a reconstructed uterine myoelectric signal = + ; Step 4: based on and , as well as and , and as well as and , and as well as and Constructing a vectorized reconstruction network Total loss function And train a vectorized reconstruction network. The trained vectorized reconstruction model is obtained. ; Step 5: and Input the vectorized reconstruction model after training In the middle, and the low-frequency encoder after training and the trained high-frequency encoder The trained discrete coding units are then processed to obtain low-frequency discrete latent representations and high-frequency discrete latent representations, which are then expanded into low-frequency discrete sequences in temporal order. With high-frequency discrete sequences ; Step 6: Construct a low-frequency bidirectional Transformer prior model and high-frequency bidirectional Transformer prior models and will and Low-frequency mask sequence after random masking and high frequency mask sequence As respectively and The training data is used to analyze... and Training was performed separately, resulting in the trained low-frequency prior models. and the high-frequency prior model after training ; Step 7: Reconstruct the model based on the trained vectorization Low-frequency prior models after training and the high-frequency prior model after training This generates the final uterine electromyography signal. .

[0007] The method for generating uterine electromyography signals based on a deep generative model described in this invention is characterized in that step 3 is performed according to the following steps: Step 3.1: The low-frequency encoder and high frequency encoder To each and performing processing, and outputting a low-frequency continuous latent representation accordingly and a high-frequency continuous latent representation ; Step 3.2: the discrete coding unit respectively maps and to a preset low-frequency codebook and a high-frequency codebook , and obtains a low-frequency discrete latent representation and a high-frequency discrete latent representation ; Step 3.3: the low-frequency decoder and the high-frequency decoder respectively perform processing on and , and output a reconstructed low-frequency component and a reconstructed high-frequency component , and respectively perform inverse short-time Fourier transform on and , and obtain and .

[0008] Further, step 4 is performed according to the following steps: Step 4.1: constructing a reconstruction loss using formula (1) : (1) Step 4.2: constructing a quantization loss using formula (2) : (2) In formula (2), indicates gradient stop; Step 4.3: constructing a commitment loss using formula (3) : (3) Step 4.4: constructing a total loss function using formula (4) : (4) In formula (4), is a weighting parameter of the commitment loss; Step 4.5: using an AdamW optimizer to perform back propagation training on the vectorization reconstruction network , and calculating the total loss function to update the network parameters until the total loss function converges, thereby obtaining a trained vectorization reconstruction model .

[0009] Furthermore, step 6 is performed as follows: Step 6.1: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require and Enter them separately and The process is performed to obtain the classification distribution of low-frequency mask positions. Classification and distribution of high-frequency mask positions ; Step 6.2: Construct the cross-entropy loss using equation (5) : (5) Step 6.3: Use the AdamW optimizer to optimize the performance of each product. and Perform backpropagation training and calculate To update network parameters until The process continues until convergence is achieved, thus obtaining the trained low-frequency prior model. and the trained low- and high-frequency prior models .

[0010] Furthermore, step 7 is performed as follows: Step 7.1: For and Perform full masking to obtain low-frequency and high-frequency full mask sequences; iterate the low-frequency channels of the low-frequency full mask sequence to obtain a low-frequency discrete sequence. ;exist Under these conditions, high-frequency channel iteration is performed on the high-frequency full-mask sequence to obtain the high-frequency discrete sequence. ; Step 7.2: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require and Mapped to the corresponding low-frequency quantization latent vector With high-frequency quantized latent vectors Then, enter the following respectively and The process is performed to generate a low-frequency time-frequency domain feature representation. Sum and high-frequency time-frequency domain feature representation ; Step 7.3: For and Perform inverse short-time Fourier transforms on each, and the resulting low-frequency time-domain signals and high-frequency time domain signals and will and After addition, the final uterine electromyography signal is obtained. .

[0011] The electronic device comprises a memory and a processor, and is characterized in that the memory is used for storing a program supporting the processor to execute the uterine myoelectric signal generation method, and the processor is configured to execute the program stored in the memory.

[0012] The computer readable storage medium stores a computer program, and the computer program is characterized in that when the computer program is run by a processor, the steps of the uterine myoelectric signal generation method are executed.

[0013] Compared with the prior art, the present application has the following specific beneficial effects: 1、The present application can capture and maintain the characteristics of different frequency bands of the signal by performing time-frequency domain decomposition on the uterine myoelectric signal and constructing a reconstruction process on the low-frequency and high-frequency branches, so that the generated uterine myoelectric signal is closer to the real data in the frequency spectrum structure; 2、The present application converts the complex continuous signal generation problem into a discrete sequence generation problem through vector quantization and token sequence modeling, which not only improves the stability of the training process, but also enhances the diversity of the generated samples and avoids the problem of single mode or mode collapse; 3、The present application uses a bidirectional Transformer prior model to capture short-range and long-range dependencies simultaneously, thereby ensuring the consistency and authenticity of the generated signal in overall trend and local details; 4、The present application can generate high-quality and reasonably distributed uterine myoelectric signals under the condition of limited samples, thereby effectively alleviating the influence of data scarcity and class imbalance on downstream tasks. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 The method flowchart of the present application; Figure 2 The uterine myoelectric signal reconstruction process schematic diagram of the present application; Figure 3 The two examples of uterine myoelectric signal waveform graphs reconstructed by the present application; Figure 4 The two examples of uterine myoelectric signal waveform graphs generated by the present application; Figure 5 The PCA visualization diagram of the reconstructed and real uterine myoelectric signals and the generated and real uterine myoelectric signals. DETAILED DESCRIPTION

[0015] In this embodiment, a uterine myoelectric signal generation method based on a deep generative model is a stable and efficient method that can balance long-term time sequence dependence modeling and time-frequency domain feature preservation to solve the problems of uterine myoelectric signal in data scarcity, class imbalance, and signal authenticity and stability, and provide more reliable data support for labor monitoring and premature birth risk prediction. Specifically, as shown in Figure 1 The method is performed according to the following steps: Step 1: Collect the uterine myoelectric signal of the pregnant woman's abdomen through the m-channel surface electrode and pre-process to obtain the pre-processed uterine myoelectric signal sequence , let the pre-processed uterine myoelectric signal of any channel in be denoted as ; m represents the number of channels; In this example, m = 4-channel electrodes E1, E2, E3, and E4 are symmetrically placed around the pregnant woman's navel, with an electrode spacing of 7 cm. The sampling rate is set to Hz, the collection time T is 30 minutes. The number of sample points N = 36000, and the gestational age of each pregnant woman is after 31 weeks. In order to avoid noise interference from the outside, the original 4-channel electrodes are pairwise differentiated to obtain three bipolar channels, where S1 = E2 − E1, S2 = E2 − E3, and S3 = E4 − E3. The three channels EHG are used as the reference for generating uterine myoelectricity, and the S2 channel is selected as the original uterine myoelectric signal. The bandwidth of the band-pass filter is set to 0.3-3Hz, and a fourth-order Butterworth filter is used for filtering and denoising. In order to eliminate the transient response of the filter, the first and last 5 minutes of data are removed, and the middle 20000-point uterine myoelectric signal sequence is normalized according to the above formula.

[0016] Step 2: Perform short-time Fourier transform on to obtain the time-frequency representation u, and perform frequency component band decomposition on u to obtain the low-frequency component and the high-frequency component ; after inverse short-time Fourier transform is performed on and , the low-frequency time-domain signal and the high-frequency time-domain signal are obtained accordingly; In this example, the STFT transform method is shown in formula (1): (1) In formula (1), is the input uterine signal sequence, is the window function, is the time n and frequency a two-dimensional function. After obtaining the time-frequency domain representation, the decomposition is performed according to the size of the frequency components: the component corresponding to the lowest frequency in the middle is selected as the low-frequency component, and the remaining frequency components are divided into high-frequency components. The window function of the short-time Fourier transform is set to the Hanning window, the window length is set to 8, the frequency axis range is set to [1, 2, 3, 4, 5], for the frequency components, the lowest frequency component is assigned to the low-frequency component, and the high-frequency component region is filled with 0, and the remaining frequency components are assigned to the high-frequency component, and the low-frequency component region is filled with 0.

[0017] Step 3: Constructing a vectorized reconstruction network , including: a low-frequency encoder and a high-frequency encoder , a discrete encoding unit, a low-frequency decoder and a high-frequency decoder , and encoding, quantization and decoding reconstruction of and to obtain reconstructed low-frequency time domain signal and reconstructed high-frequency time domain signal , and composing a reconstructed uterine electromyography signal = + ; in this example, the reconstruction process is shown in Figure 2 .

[0018] Step 3.1: Low-frequency encoder and high-frequency encoder of respectively process and , and output low-frequency continuous latent representation and high-frequency continuous latent representation ; The number of low-frequency and high-frequency encoder convolution blocks is set to 9, i.e. the downsampling rate is 2 9 =512, and the number of residual blocks is set to 4 to stabilize the training process. The convolution block is composed of one two-dimensional convolution layer, one two-dimensional batch normalization (BatchNorm) layer and one ReLU activation function layer, the convolution kernel size is 3x4, the step is (1, 2), and the padding is (1, 1). The residual block is a two-dimensional convolution layer, the convolution kernel size is 3x3, the step is (1, 1), and the padding is (1, 1). The initial dimension of the encoder is set to 4, and the hidden layer dimension is set to 64.

[0019] Step 3.2: The discrete encoding unit respectively maps and to the preset low-frequency codebook and high-frequency codebook , and correspondingly obtains low-frequency discrete latent representation ​High-frequency discrete latent representation ; In this example, considering the complex patterns of uterine electromyography signals, the codebook size for both low-frequency and high-frequency components is set to 64 during vectorization, the same as the encoder's hidden layer dimension.

[0020] Step 3.3: Low-frequency decoder and high frequency decoder To each and The process is performed, and the reconstructed low-frequency components are output accordingly. With the reconstructed high-frequency components and respectively and Performing the inverse short-time Fourier transform yields the following results: and ; In this example, the number of upsampled convolutional blocks and residual blocks in the decoder is the same as the encoder settings in step 3.1. The convolutional structure of the decoder is transposed convolution, which mainly serves as upsampling.

[0021] Step 4: Based on and as well as and , and as well as and , and as well as and Constructing a vectorized reconstruction network Total loss function And train a vectorized reconstruction network. The trained vectorized reconstruction model is obtained. ; Step 4.1: Construct the reconstruction loss using equation (1) : (1) Step 4.2: Construct the quantified loss using equation (2) : (2) In equation (2), This indicates gradient stopping; Step 4.3: Construct the commitment loss using equation (3) : (3) Step 4.4: Construct the total loss function using equation (4) : (4) In equation (4), Weighted parameters for committed losses; Step 4.5: Reconstruct the vectorized network using the AdamW optimizer Perform backpropagation training and calculate the total loss function. To update network parameters until the total loss function is reached. The process continues until convergence is achieved, thus obtaining the trained vectorized reconstruction model. In this example, the AdamW optimizer has an initial learning rate of 1e-3, with 4 samples input per batch and 50,000 training iterations.

[0022] Step 5: and Input the vectorized reconstruction model after training In the middle, and the low-frequency encoder after training and the trained high-frequency encoder The discrete coding unit processes the data to obtain low-frequency discrete latent representations and high-frequency discrete latent representations, which are then expanded into low-frequency discrete sequences in temporal order. With high-frequency discrete sequences ; In this example, the low-frequency and high-frequency bidirectional Transformer models have the same model parameters: the hidden layer dimension is 128, the number of Transformer models is 2, the self-attention head in the Transformer is 2, and the feedforward ratio is the ratio of the hidden layer dimension to the input dimension, which is set to 1 here.

[0023] Step 6: Construct a low-frequency bidirectional transformer prior model and high-frequency bidirectional transformer prior models and will and Low-frequency mask sequence after random masking and high frequency mask sequence As respectively and The training data is used to analyze... and Training was performed separately, resulting in the trained low-frequency prior models. and the high-frequency prior model after training ; Step 6.1: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require and Enter them separately and The low-frequency mask position classification distribution is obtained by processing the low-frequency mask position classification distribution and the low-frequency mask position classification distribution ; In this example, and are time-aligned and divided into training segments of the same length, position encoding is added, and random masking is performed with a masking ratio of 10%, obtaining and as training data.

[0024] Step 6.2: Construct cross-entropy loss using formula (5) : (5) Step 6.3: Use the AdamW optimizer to perform backpropagation training on and respectively, and calculate to update the network parameters until converges, thereby obtaining the trained low-frequency prior model and the trained low-high frequency prior model ; In this example, the prior modeling of the high-frequency component is performed after the prior modeling of the low-frequency component, so the cross-entropy loss of the high-frequency component is calculated under the condition of low frequency. The initial learning rate of the AdamW optimizer is 1e-3, 8 samples are input per batch, and the training number is 50000.

[0025] Step 7: Based on the trained vectorization reconstruction model , the trained low-frequency prior model , and the trained high-frequency prior model , generate the final uterine myoelectric signal signal ; Step 7.1: Perform full masking on and to obtain low-frequency full mask sequences and high-frequency full mask sequences; perform low-frequency channel iteration on the low-frequency full mask sequence to obtain low-frequency discrete sequence ; under the condition of , perform high-frequency channel iteration on the high-frequency full mask sequence to obtain high-frequency discrete sequence ; In this example, the full mask sequence is used as the starting point. First, do T rounds of iteration in the low-frequency channel: the t-th round outputs the distribution of all masked positions, and samples from it with temperature τ(t), fills according to the mask scheduling function (preferably cosine schedule), to obtain ; then ​The same T-round iteration is performed on the high-frequency channel to obtain For example, T is set to 10.

[0026] Step 7.2: Map and to the corresponding low-frequency quantized latent vectors and high-frequency quantized latent vectors respectively, and input and into the processing unit for processing, to generate low-frequency time-frequency domain feature representations and high-frequency time-frequency domain feature representations respectively; Step 7.3: Perform inverse short-time Fourier transform on and respectively, to generate low-frequency time domain signals and high-frequency time domain signals respectively, and add and to obtain the final generated uterine electromyography signal .

[0027] In this embodiment, an electronic device includes a memory for storing a program supporting a processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0028] In this embodiment, a computer readable storage medium has a computer program stored thereon, and the computer program is executed by a processor to perform the steps of the above method.

[0029] To evaluate the performance of the present application in the reconstruction and generation stages, in this embodiment, the representative visual and quantitative results are given in Figure 3 — Figure 5 . Figure 3 Two examples of original uterine electromyography (EHG) signals and their corresponding reconstructed sequences are compared; Figure 4 Two examples of EHG signals generated by the model are shown. The above images are arranged in a three-row structure of “low-frequency component—high-frequency component—low / high-frequency superimposed signal” to facilitate observation of the preservation and fusion effects of different frequency band information. In addition, to depict the differences in data distribution, this paper visualizes and compares a group of generated samples, and draws principal component analysis (PCA) dimension reduction diagrams of “reconstruction-original” and “generation-original” respectively (see Figure 5), from the embedding space perspective, to show the closeness of the distribution of both. Further, the present application adopts Fréchet Inception Distance (FID) as a distribution consistency evaluation index to measure the distribution difference between the generated samples and the real samples, and obtains a result of FID = 0.0406; the smaller the index value, the closer the two distributions, so the result shows that the EHG signals generated by the present application are consistent with the real samples in statistical distribution characteristics, and can meet the requirements of high-quality signal generation.

Claims

1. A method for uterine myoelectric signal generation based on deep generative model, characterized in that, is performed according to the following steps: Step 1: Collect the uterine electromyographic signals of the abdomen of a pregnant woman through surface electrodes of m channels and pre-process to obtain a pre-processed uterine electromyographic signal sequence , let Any channel of the pre-processed uterine electromyographic signals in ; m represents the number of channels; Step 2: Perform a short-time Fourier transform on to obtain a time-frequency representation u, and perform a frequency- dependent band decomposition of u to obtain a low-frequency component and a high-frequency component ; perform an inverse short-time Fourier transform on and to obtain a low-frequency time-domain signal and a high-frequency time-domain signal ; Step 3: Construct a vectorized reconstruction network Including: low-frequency encoders and high frequency encoder Discrete coding unit (low-frequency codebook) With high-frequency codebook ), low-frequency decoder and high frequency decoder and to and Encoding, quantization, and decoding are performed to reconstruct the low-frequency time-domain signal. With the reconstructed high-frequency time-domain signal And form the reconstructed uterine electromyographic signal = + ; Step 4: based on With And With , With And With , With And With Constructing a vectorization reconstruction network The total loss function of And train the vectorization reconstruction network Get the trained vectorization reconstruction model ; Step 5: and Input the vectorized reconstruction model after training In the middle, and the low-frequency encoder after training and the trained high-frequency encoder The trained discrete coding units are then processed to obtain low-frequency discrete latent representations and high-frequency discrete latent representations, which are then expanded into low-frequency discrete sequences in temporal order. With high-frequency discrete sequences ; Step 6: Constructing the low-frequency bidirectional Transformer prior model and the high-frequency bidirectional Transformer prior model , and and the low-frequency mask sequence after random masking and the high-frequency mask sequence respectively as and training data for training and respectively, and obtaining the trained low-frequency prior model and the trained high-frequency prior model respectively; Step 7: Generating final uterine electromyography signal based on the trained vectorization reconstruction model , trained low frequency prior model , and trained high frequency prior model .​ 2. The method of claim 1, wherein, Step 3 is performed according to the following steps: Step 3.1: The low frequency encoder and the high frequency encoder of respectively and are processed, outputting a low frequency continuous latent representation and a high frequency continuous latent representation respectively Step 3.2: The discrete coding units respectively... and Mapped to a preset low-frequency codebook With high-frequency codebook In this process, the low-frequency discrete latent representation is obtained. High-frequency discrete latent representation ; Step 3.3: the low-frequency decoder and the high-frequency decoder processes the low-frequency component and the high-frequency component respectively, and outputs the reconstructed low-frequency component and the reconstructed high-frequency component respectively, and performs inverse short-time Fourier transform on the low-frequency component and the high-frequency component respectively, and obtains the low-frequency component and the high-frequency component respectively.

3. The method of claim 2, wherein the method is based on a deep generative model. Step 4 is performed according to the following steps: Step 4.1 : Constructing the reconstruction loss with formula (1) : (1) Step 4.2: Constructing the quantization loss with formula (2) : (2) In formula (2), indicates a gradient stop; Step 4.3: Constructing the commitment loss with formula (3) : (3) Step 4.4: Constructing the total loss function with formula (4) : (4) In formula (4), is a weighted parameter of commitment loss; Step 4.5: Vectorization reconstruction network using AdamW optimizer Backpropagation training is performed and the total loss function is calculated The network parameters are updated until the total loss function converges, resulting in a trained vectorization reconstruction model .

4. The method of claim 3, wherein the method is based on a deep generative model. Step 6 is performed according to the following steps: Step 6.1: inputting the low frequency mask positions and the high frequency mask positions into the low frequency mask position classification distribution model and the high frequency mask position classification distribution model respectively, to obtain the low frequency mask position classification distribution and the high frequency mask position classification distribution respectively. Step 6.2: Constructing the cross-entropy loss with formula (5) : (5) Step 6.3: Train the low-frequency prior model and the low-high frequency prior model using the AdamW optimizer, respectively, and calculate the gradient to update the network parameters until convergence, thereby obtaining the trained low-frequency prior model and the trained low-high frequency prior model. and and .​​​ 5. The method of claim 4, wherein the method is based on a deep generative model. Step 7 is performed according to the following steps: Step 7.1: Perform full masking on and to obtain a low-frequency full masking sequence and a high-frequency full masking sequence; perform low-frequency channel iteration on the low-frequency full masking sequence to obtain a low-frequency discrete sequence ; under the condition that , perform high-frequency channel iteration on the high-frequency full masking sequence to obtain a high-frequency discrete sequence ; Step 7.2: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] and Mapped to the corresponding low-frequency quantization latent vector With high-frequency quantized latent vectors Then, enter the following respectively and The process is performed to generate a low-frequency time-frequency domain feature representation. Sum and high-frequency time-frequency domain feature representation ; Step 7.3: inverse short-time Fourier transform is performed on and respectively, to generate corresponding low-frequency time-domain signals and high-frequency time-domain signals respectively, and and are added to obtain the final generated uterine myoelectricity signal .

6. An electronic device comprising a memory and a processor, characterized in that The memory is configured to store a program supporting the processor to execute the method of generating a uterine myoelectric signal according to any one of claims 1-5, and the processor is configured to execute the program stored in the memory.

7. A computer-readable storage medium having stored thereon a computer program, characterized in that The computer program, when executed by the processor, performs the steps of the method of generating a uterine myoelectric signal according to any one of claims 1-5.

Images