Intelligent recognition and detection method and system for pelvic floor muscle electrical signals

Abstract

The embodiment of the present application relates to the functional detection technical field, disclose a kind of intelligent identification detection method of pelvic floor muscle electrical signal, comprising: receiving the pelvic surface electromyogram signal detected by surface electromyogram instrument, the pelvic surface electromyogram signal continuously collected is divided into corresponding short period signal segment by sliding window;Each short period signal segment is converted into myoelectric frequency domain feature by Fourier transform or wavelet change to time domain signal, and corresponding hash feature code is generated according to myoelectric frequency domain feature and hash algorithm;The similarity between hash feature code and each hash feature in hash feature library is determined by similarity matching of generated hash feature code and each hash feature in hash feature library.The intelligent identification detection method of pelvic floor muscle electrical signal in the embodiment of the present application greatly reduces the risk that effective signal is removed as noise, and assists in improving the accuracy of subsequent data identification.

Description

Technical Field

[0001] This invention relates to the field of functional testing technology, specifically to an intelligent identification and detection method and system for pelvic floor electromyography signals. Background Technology

[0002] Currently, pelvic floor dysfunction (PFD) is a common gynecological and urinary system disease, mainly including urinary incontinence and pelvic organ prolapse (such as uterine prolapse, bladder prolapse, and rectal prolapse). Its incidence increases significantly with age, severely impacting patients' quality of life. In recent years, artificial intelligence (AI) technology, with its powerful data processing and pattern recognition capabilities, has shown great potential in the field of medical diagnosis. Applying AI to the diagnosis of pelvic floor dysfunction can effectively improve diagnostic efficiency, alleviate the strain on medical resources, and facilitate early screening and intervention of the disease.

[0003] However, current artificial intelligence technologies applied to the diagnosis of pelvic floor dysfunction still have significant limitations. Existing technologies typically rely on single-modal imaging data, such as single ultrasound images or MRI images, for auxiliary analysis. This still presents challenges in terms of accuracy. Summary of the Invention

[0004] To address the aforementioned shortcomings, this invention discloses an intelligent identification and detection method for pelvic floor electromyography (EMG) signals, which can detect weak EMG signals and greatly reduce the possibility of removing effective signals as noise.

[0005] The first aspect of this invention discloses an intelligent recognition and detection method for pelvic floor electromyography signals, comprising:

[0006] It receives pelvic floor surface electromyography signals detected by a surface electromyography instrument, and divides the continuously acquired pelvic floor surface electromyography signals into corresponding short-term signal segments through a sliding window;

[0007] For each short-time signal segment, the time-domain signal is converted into electromyographic frequency domain features through Fourier transform or wavelet transform, and a corresponding hash feature code is generated based on the electromyographic frequency domain features and a hash algorithm.

[0008] The generated hash feature code is matched with each hash feature in the hash feature library to determine the similarity between the hash feature code and each hash feature in the hash feature library; if the similarity between the two is within a first set interval, the corresponding short-term signal segment is determined to be a valid signal segment; if the similarity between the two is within a second set interval, the corresponding short-term signal is determined to be a signal segment to be confirmed; if the similarity between the two is within a third set interval, the corresponding short-term signal is determined to be an invalid signal.

[0009] Acquire surface electromyography (EMG) signals within a predetermined time range before and after the signal segment to be confirmed, and input the acquired EMG signals into a signal recognition model for identification to determine whether the signal segment to be confirmed is a valid signal.

[0010] As an optional implementation, in the first aspect of the present invention, the step of acquiring surface electromyography (EMG) signals within a predetermined time range before and after the signal segment to be confirmed, and inputting the acquired EMG signals into a signal recognition model for identification to determine whether the signal segment to be confirmed is a valid signal, includes: acquiring surface EMG signals within a predetermined time range before and after the signal segment to be confirmed;

[0011] The time-domain feature extraction module extracts time-domain features from the surface electromyography (EMG) signals within a set time range to obtain the corresponding time-domain features, and the frequency-domain feature extraction module extracts frequency-domain features from the surface EMG signals within a set time range to obtain the frequency-domain features.

[0012] The time-domain and frequency-domain features are flattened and concatenated to form a joint feature vector. The hash features in the second set interval are encoded into a dense vector through an embedding matrix. The joint feature vector and the dense vector are input into the attention layer to obtain the enhanced composite features.

[0013] The enhanced composite features are input into a bidirectional LSTM network to obtain the dynamic changes of surface electromyography signals over time. The classification decision layer outputs the probability of valid signals to determine whether the signal segment to be confirmed is a valid signal.

[0014] As an optional implementation, in the first aspect of the present invention, before segmenting the continuously acquired pelvic floor surface electromyography signals into corresponding short-term signal segments via a sliding window, the method further includes:

[0015] Determine the initial window parameters of the sliding window, including the sampling frequency, the initial window duration, and the initial overlap rate;

[0016] Calculate the sample entropy of the signal within the current window. If the current sample entropy is greater than a first set value, reduce the window duration of the sliding window. If the current sample entropy is less than a second set value, increase the window duration of the sliding window. If the current sample entropy is between the first and second set values, maintain the initial window parameters.

[0017] When the sample entropy changes, the window duration is adjusted only when the sample entropy of multiple consecutive sliding windows changes in the same way.

[0018] As an optional implementation, in the first aspect of the present invention, the step of converting the time-domain signal into electromyographic frequency domain features for each short-time signal segment through Fourier transform or wavelet transform, and generating a corresponding hash feature code based on the electromyographic frequency domain features and a hash algorithm, includes:

[0019] The short-segment signal is decomposed into at least six sub-bands using wavelet packet transform;

[0020] For each sub-band, the corresponding energy entropy and band energy operator are calculated, wherein the band energy operator estimates the instantaneous energy of the signal by calculating the second difference of the signal;

[0021] The energy entropy and band energy operator of each sub-band are concatenated to form a high-dimensional feature vector;

[0022] A set number of hash functions are generated using random projection locality-sensitive hashing. The output of each hash function is calculated on the high-dimensional feature vector, and the set number of hash outputs are concatenated in order to form the corresponding hash feature code.

[0023] As an optional implementation, in the first aspect of the present invention, the step of performing similarity matching between the generated hash feature code and each hash feature in the hash feature library to determine the similarity between the hash feature code and each hash feature in the hash feature library includes:

[0024] Obtain the generated hash feature encoding, traverse all hash features in the hash feature library, and calculate the Hamming distance between them;

[0025] If the Hamming distance between the two is less than the first set value, then the corresponding hash feature is used as a candidate feature;

[0026] Construct a cost matrix between candidate features and each hash feature in the hash feature library, and dynamically program the filling of the cumulative cost matrix to obtain the dynamic time warped distance;

[0027] The similarity between the hash feature encoding and each hash feature in the hash feature library is determined based on the dynamic time rule distance.

[0028] As an optional implementation, in the first aspect of the present invention, the intelligent identification and detection method further includes:

[0029] When a valid signal is detected, the hash value of the corresponding valid signal is extracted and added to a temporary buffer. When the number of valid signals in the temporary buffer meets the requirements, the next step is executed.

[0030] Run the DBSCAN algorithm on all historical unclustered hash values ​​in the buffer, where the DBSCAN algorithm includes the neighborhood radius and the minimum number of samples;

[0031] If the Hamming distance between the hash of a valid signal and the centroid of a cluster in the feature library is not greater than the neighborhood radius, then the new signal is determined to be assigned to that cluster, and the cluster size and last update time are updated.

[0032] If the hash of the valid signal and the Hamming distance between the centroids of each cluster in the feature library are both greater than the neighborhood radius, and the number of neighborhood samples of the cluster containing the corresponding valid signal is greater than a set number, then a new cluster is created.

[0033] As an optional implementation, in the first aspect of the present invention, the intelligent identification and detection method further includes:

[0034] All valid signal segments are spliced ​​together according to time series and then input into the pelvic floor muscle resistance detection model for identification to identify whether there are abnormal segments.

[0035] The training data in the pelvic floor muscle resistance detection model includes positive sample data and negative sample data, wherein the negative sample data includes noise signals and non-target muscle contraction signals.

[0036] A second aspect of this invention discloses an intelligent recognition and detection system for pelvic floor electromyography signals, comprising:

[0037] Receiver module: Used to receive pelvic floor surface electromyography signals detected by a surface electromyography instrument, and to divide the continuously acquired pelvic floor surface electromyography signals into corresponding short-term signal segments through a sliding window;

[0038] Transformation module: used to convert the time-domain signal into electromyographic frequency domain features for each short-time signal segment through Fourier transform or wavelet transform, and generate corresponding hash feature codes based on the electromyographic frequency domain features and hash algorithm;

[0039] Matching module: This module performs similarity matching between the generated hash feature code and each hash feature in the hash feature library to determine the similarity between the hash feature code and each hash feature in the hash feature library. If the similarity between the two is within a first set interval, the corresponding short-term signal segment is determined to be a valid signal segment. If the similarity between the two is within a second set interval, the corresponding short-term signal is determined to be a signal segment to be confirmed. If the similarity between the two is within a third set interval, the corresponding short-term signal is determined to be an invalid signal.

[0040] The identification module is used to acquire surface electromyography (EMG) signals within a set time range before and after the signal segment to be confirmed, and input the acquired EMG signals into the signal identification model for identification to determine whether the signal segment to be confirmed is a valid signal.

[0041] A third aspect of the present invention discloses an electronic device, comprising: a memory storing executable program code; a processor coupled to the memory; the processor calling the executable program code stored in the memory to execute the intelligent recognition and detection method for pelvic floor electromyography signals disclosed in the first aspect of the present invention.

[0042] A fourth aspect of the present invention discloses a computer-readable storage medium storing a computer program, wherein the computer program causes a computer to execute the intelligent recognition and detection method for pelvic floor electromyography signals disclosed in the first aspect of the present invention.

[0043] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0044] The intelligent identification and detection method for pelvic floor electromyography signals in this embodiment of the invention achieves comprehensive identification of pelvic floor electromyography signals by using a combination of short-time segmentation, frequency domain feature extraction, hash coding, and context model recognition. This greatly reduces the risk of effective signals being removed as noise and helps improve the accuracy of subsequent data identification. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart illustrating the intelligent recognition and detection method for pelvic floor electromyography signals disclosed in an embodiment of the present invention.

[0047] Figure 2 This is a schematic diagram of the process for effective identification of surface electromyography signals disclosed in an embodiment of the present invention;

[0048] Figure 3 This is a schematic diagram of the hash feature encoding generation process disclosed in the embodiments of the present invention;

[0049] Figure 4 This is a schematic diagram of the structure of an intelligent recognition and detection system for pelvic floor electromyography signals provided in an embodiment of the present invention;

[0050] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] It should be noted that the terms "first," "second," "third," "fourth," etc., in the specification and claims of this invention are used to distinguish different objects, not to describe a specific order. The terms "comprising" and "having," and any variations thereof, in the embodiments of this invention are intended to cover non-exclusive inclusion. Exemplarily, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or devices.

[0053] Example 1

[0054] Please see Figure 1 , Figure 1 This is a flowchart illustrating the intelligent recognition and detection method for pelvic floor electromyography signals disclosed in this invention. The execution entity of the method described in this invention is an execution entity composed of software and / or hardware. This execution entity can receive relevant information via wired or / or wireless means and can send certain instructions. It may also have certain processing and storage functions. This execution entity can control multiple devices, such as remote physical servers or cloud servers and related software, or local hosts or servers and related software that perform related operations on devices located in a certain place. In some scenarios, multiple storage devices can also be controlled; these storage devices may be placed in the same location as the devices or in different locations. Figure 1 As shown, the intelligent recognition and detection method based on pelvic floor electromyography signals includes the following steps:

[0055] S101: Receives pelvic floor surface electromyography signals detected by a surface electromyography instrument, and divides the continuously acquired pelvic floor surface electromyography signals into corresponding short-term signal segments through a sliding window;

[0056] S102: For each short-time signal segment, the time-domain signal is converted into electromyographic frequency domain features through Fourier transform or wavelet transform, and a corresponding hash feature code is generated based on the electromyographic frequency domain features and the hash algorithm;

[0057] S103: The generated hash feature code is matched with each hash feature in the hash feature library to determine the similarity between the hash feature code and each hash feature in the hash feature library; if the similarity between the two is within a first set interval, the corresponding short-term signal segment is determined to be a valid signal segment; if the similarity between the two is within a second set interval, the corresponding short-term signal is determined to be a signal segment to be confirmed; if the similarity between the two is within a third set interval, the corresponding short-term signal is determined to be an invalid signal.

[0058] S104: Acquire surface electromyography (EMG) signals within a set time range before and after the signal segment to be confirmed, and input the acquired EMG signals into a signal recognition model for identification to determine whether the signal segment to be confirmed is a valid signal.

[0059] In traditional long-term signal analysis, minute signals are easily overwhelmed by overall noise, while short-term segmentation can focus on the instantaneous changes in the signal, avoiding its being ignored due to low signal strength. The overlapping design of the sliding window ensures that minute signals are not truncated by the window boundaries, guaranteeing feature integrity.

[0060] This invention uses a sliding window to segment continuous pelvic floor surface electromyography (EMG) signals into short-term signal segments, which can capture local features of the signals, adapt to the non-stationary characteristics of EMG signals, facilitate subsequent analysis of different features, and reduce computational load while improving processing efficiency.

[0061] Specifically, Fourier transform converts weak time-domain signals into frequency-domain energy distributions, enabling the identification of minute energy peaks within specific frequency bands, such as the characteristic frequency bands during pelvic floor muscle contraction, thus avoiding interference from time-domain noise on signal amplitude. Wavelet transform's multi-resolution analysis is more suitable for non-stationary electromyographic signals, capturing the time-frequency characteristics of minute signals at different scales, such as short-duration high-frequency pulses, and exhibiting stronger robustness to signals with low signal-to-noise ratios.

[0062] The hash algorithm provided in this embodiment compresses high-dimensional frequency domain features into binary codes, preserving the core frequency domain features of minute signals through feature binarization, thus avoiding feature loss due to low signal amplitude. More specifically, the fault-tolerant design of the hash code can tolerate amplitude fluctuations of minute signals, ensuring that similar features (such as different intensities of micro-contractions of the same muscle) are grouped together, reducing the false deletion of valid signals. Generating hash feature codes based on electromyography (EMG) frequency domain features and the hash algorithm transforms complex frequency domain features into a concise encoding form, facilitating storage and comparison. Simultaneously, the hash algorithm possesses good anti-interference and uniqueness, contributing to improved accuracy and efficiency of feature matching. The hash code, generated based on EMG frequency domain features and the hash algorithm, is efficient and robust. It can map high-dimensional frequency domain features into compact binary codes, significantly reducing data storage space and computational costs, while maintaining the similarity of signal features; that is, hash codes generated from similar EMG frequency domain features are also relatively similar, facilitating subsequent similarity matching.

[0063] In addition to the advantages mentioned above, the solution of this invention also sets up a three-level matching method to achieve efficient data filtering. The first set interval (high similarity) directly identifies known effective signal patterns (such as the frequency domain characteristics of standard pelvic floor muscle contraction), ensuring that strong signals pass through quickly. The second set interval (medium similarity) classifies suspected small effective signals into the pending confirmation category, avoiding misjudgment as noise due to low signal amplitude and insufficient matching with the feature library. For example, the frequency domain characteristics of small electromyographic signals may differ from standard features, but the core frequency components still exist, and the medium similarity interval can retain such signals. The third set interval (low similarity) filters strong noise, but through the buffering of the first two intervals, the probability of small signals being misjudged is reduced.

[0064] More specifically, during implementation, because suspected noise can vary significantly, it's necessary to further verify the validity of the signal segment to be confirmed. In practice, signals within a preceding and following time window (e.g., 500ms before and after the signal) can be extracted and input into a deep learning model (e.g., LSTM / CNN) to analyze temporal correlations. Small, valid signals usually have physiological correlations with preceding and following signals (e.g., the gradual change in electromyographic signals), while noise is mostly random fluctuation. The model can learn the temporal patterns of muscle activity (e.g., pre-contraction electromyographic warm-up and post-contraction decay) to determine whether the signal to be confirmed conforms to physiological logic, avoiding false deletions caused by isolated judgments. Combining these preceding and following correlations significantly improves the accuracy of the judgment results.

[0065] For example, for a short-duration signal with low amplitude, if a similar sequence of weak signals appears before and after it, the model can infer that it represents continuous, subtle electromyographic activity rather than noise, thus preserving the effective signal. The method of this invention is particularly suitable for scenarios such as postpartum pelvic floor muscle function assessment and weak electromyographic monitoring in patients with urinary incontinence. It can capture muscle micro-contraction signals that are difficult to identify using traditional methods, providing more accurate data support for clinical diagnosis. For instance, in postpartum pelvic floor muscle rehabilitation, the identification of subtle signals can help determine early signs of muscle nerve recovery, avoiding misjudgments of rehabilitation progress due to weak signals.

[0066] More preferably, such as Figure 2 As shown, the step of acquiring surface electromyography (EMG) signals within a predetermined time range before and after the signal segment to be confirmed, and inputting the acquired EMG signals into a signal recognition model for identification to determine whether the signal segment to be confirmed is a valid signal, includes:

[0067] S1041: Acquire surface electromyography signals within a set time range before and after the signal end to be confirmed;

[0068] S1042: The time-domain feature extraction module extracts time-domain features from the surface electromyography (EMG) signal within a set time range to obtain the corresponding time-domain features, and the frequency-domain feature extraction module extracts frequency-domain features from the surface EMG signal within a set time range to obtain the frequency-domain features.

[0069] S1043: Flatten the time-domain features and frequency-domain features and concatenate them to form a joint feature vector. Use an embedding matrix to encode the hash features in the second set interval into a dense vector. Input the joint feature vector and the dense vector into the attention layer to obtain the enhanced composite features.

[0070] S1044: The enhanced composite features are input into a bidirectional LSTM network to obtain the dynamic change information of the surface electromyography signal in the time series. The classification decision layer outputs the probability of the effective signal to determine whether the signal segment to be confirmed is a valid signal.

[0071] Specifically, acquire surface electromyography (EMG) signals for T seconds before and after the signal segment to be confirmed (e.g., T = 0.5 seconds, corresponding to 1000 time points at 1000Hz sampling) to form a time window of 2T seconds in length, ensuring that the contextual information of the signal is included (e.g., the baseline before contraction and the recovery process after contraction).

[0072] Specifically, the temporal feature extraction module calculates parameters such as peak value, root mean square (RMS), integrated electromyography (iEMG), and zero-crossing rate, reflecting the intensity and duration of muscle contraction. Implementation methods include calculation via a sliding window (window length 200ms, stride 50ms) or automatic extraction via a 1D convolutional network (e.g., using a convolutional layer with kernel_size=11 to capture local temporal patterns).

[0073] Frequency domain feature extraction module: This module obtains the power spectral density (PSD) of the signal through short-time Fourier transform (STFT, window length 256, step size 64) or wavelet transform (e.g., db4 wavelet decomposition to 8 levels). Specific feature parameters include average power frequency (MPF), median frequency (MF), and high-frequency energy proportion (>150Hz), used to distinguish between fast and slow muscle fiber activity. A joint feature vector is formed by concatenating time-domain and frequency-domain features, thus obtaining multi-dimensional feature data.

[0074] In specific implementation, the solution of this invention can also perform hash feature processing and attention fusion: the 64-bit hash code is converted into a 128-dimensional dense vector through the embedding matrix Embedding(64,128), enabling the discrete hash value to have the semantic expression capability of continuous space. The attention layer implementation logic takes the following inputs: joint feature vector + hash dense vector, concatenated into multi-dimensional features. Attention calculation: weights are generated using nn.Linear(568,1), and after Sigmoid activation, weight values ​​between 0 and 1 are obtained, used for weighted fusion of joint features and hash features.

[0075] weighted_features=joint_features*attn_weight.

[0076] This invention also provides bidirectional LSTM and timing modeling.

[0077] Input dimension: Enhanced composite features, adjusted to [time step, batch size, feature dimension]. Network structure: 2-layer bidirectional LSTM, 128 hidden units per layer, output dimension [time step, batch size, 256] (bidirectional concatenation). Classification decision layer input: Utilizing the output of the last time step of the LSTM (capturing the overall trend of the signal) or the result of global average pooling (fusing information from all time steps), input to the fully connected layer for binary classification (valid signal / invalid signal). This method can fully realize the identification of the validity of the corresponding signal.

[0078] Specifically, in this embodiment of the invention, signals within a set time window (e.g., ±500ms) before and after the signal segment to be confirmed are extracted to construct a signal context, avoiding misjudgments caused by isolated analysis. Small, effective signals (such as micro-contractions of the pelvic floor muscles) are often accompanied by temporal continuity between preceding and following signals (e.g., gradual changes in electromyographic amplitude), while noise is mostly random pulses. Contextual analysis can capture this physiological correlation. For example, if the signal segment to be confirmed is a weak electromyographic peak, there may be an electromyographic warm-up (slow increase in amplitude) in the first 200ms, followed by a decay process in the last 300ms. This temporal characteristic can serve as a key basis for judging validity.

[0079] Specifically, time-domain features such as root mean square (RMS), zero-crossing rate (ZC), and waveform length (WL) are extracted to describe the amplitude changes and waveform morphology of the signal. For example, an effective signal of pelvic floor muscle contraction often exhibits a curve where the RMS gradually increases and then decreases, while the RMS of noise may suddenly jump and be irregular. For small signals, time-domain features can reflect their weak but continuous amplitude changes, avoiding misjudgment as noise due to low amplitude.

[0080] By flattening and concatenating time-domain features (such as RMS and ZC) and frequency-domain features (such as PSD and FC), a multi-dimensional feature vector containing the signal's "amplitude-frequency-time sequence" is formed, avoiding the loss of information from single-dimensional features. For example, a small signal may have low amplitude in the time domain but feature peaks in the frequency domain; combining features can integrate information from both to improve the recognition rate. The hash feature encoding (binary vector) is converted into a dense real-valued vector through an embedding matrix, preserving the "moderate similarity" feature information from the initial classification (such as partial matching patterns with known valid signals). This operation combines the "coarse-grained" features of the hash encoding with the "fine-grained" time-frequency features of the context, forming a complementary pair.

[0081] The attention mechanism automatically learns the feature dimensions that are more important for determining the validity of a signal from the joint feature vector and the hash-dense vector. For example, for small signals, the attention layer may assign higher weights to frequency domain features (such as the energy of characteristic frequency bands) to suppress interference from temporal noise; for signals close to noise, it enhances the focus on temporal continuity features. Through adaptive weight allocation, the discriminative power of small signal features is improved, and the missed detection of valid signals due to feature dimension redundancy is reduced.

[0082] Bidirectional LSTM networks can simultaneously learn the forward (past to present) and backward (future to present) temporal information of signals, modeling the physiological patterns of pelvic floor electromyography (EMG) activity (such as the complete contraction-relaxation process). Even small, effective signals with low amplitude in a single time period can be captured by LSTM through their "gradual change process" over time (such as a slow increase in EMG in the first period and a slow decrease in the second period), while noise does not exhibit this pattern. For example, a weak signal lasting 200ms can be inferred by LSTM to be a complete small EMG contraction if similar weak signals exist in the preceding and following periods and exhibit a "increase-peak-decrease" temporal pattern. The probability of an effective signal is output through fully connected layers and a softmax function, enabling quantitative judgment of the signal. For example, a probability threshold of 0.7 can be set: a probability ≥ 0.7 is considered an effective signal, 0.3-0.7 is considered pending confirmation (requiring further manual review), and < 0.3 is considered noise. This probabilistic decision-making is more flexible than hard classification, and the threshold can be adjusted to balance "missed" and "false" judgments.

[0083] More preferably, before segmenting the continuously acquired pelvic floor surface electromyography signals into corresponding short-term signal segments via a sliding window, the method further includes:

[0084] Determine the initial window parameters of the sliding window, including the sampling frequency, the initial window duration, and the initial overlap rate;

[0085] Calculate the sample entropy of the signal within the current window. If the current sample entropy is greater than a first set value, reduce the window duration of the sliding window. If the current sample entropy is less than a second set value, increase the window duration of the sliding window. If the current sample entropy is between the first and second set values, maintain the initial window parameters.

[0086] When the sample entropy changes, the window duration is adjusted only when the sample entropy of multiple consecutive sliding windows changes in the same way.

[0087] In this invention, sample entropy is used to measure the complexity and irregularity of a signal. In pelvic floor electromyography (EMG) signals, effective signals (such as muscle contractions) exhibit a certain regularity due to physiological activity (such as contraction-relaxation waveforms), resulting in lower sample entropy; while noise (such as electrode interference) or minute signals (such as weak EMG signals) may have higher sample entropy due to low amplitude or variable morphology.

[0088] By calculating the sample entropy of the signal within a window, the "regularity" of the current signal segment can be assessed in real time, providing a quantitative basis for adjusting window parameters and avoiding the insufficient adaptability of a fixed window to signals with different characteristics. When the sample entropy is high, the signal may contain small effective signals or high-frequency noise. Shortening the window duration (e.g., from 500ms to 200ms) can divide the signal into shorter segments, preventing small signals within a long window from being averaged out by noise.

[0089] For example, if a signal contains multiple short, weak electromyographic pulses (intervals of 100ms), a long window will blur it into a low-amplitude fluctuation, while a short window can capture the characteristics of each pulse individually, improving the resolution of small signals.

[0090] When the sample entropy is low, the signal may be a regular effective signal or low-frequency noise. Extending the window duration (e.g., from 200ms to 800ms) can accumulate more signal energy, enhance the ability to capture weak but continuous effective signals (e.g., continuous micro-contraction of pelvic floor muscles), and avoid incomplete signal segments due to an excessively short window.

[0091] When the sample entropy is moderate, the signal complexity is within a reasonable range. At this time, maintaining the initial parameters can avoid the computational overhead caused by frequent window adjustments, while ensuring the stability of signal analysis.

[0092] Specifically, the window duration is only adjusted when the sample entropy of multiple consecutive windows changes by the same amount. This filters out sample entropy fluctuations caused by instantaneous noise and prevents frequent fluctuations in window parameters. For example, if a window experiences a momentary high sample entropy due to poor electrode contact, single-window adjustment might misinterpret it as a minor signal and shorten the window. However, continuous multi-window verification allows for confirmation from subsequent windows. If the sample entropy of subsequent windows returns to normal, no adjustment is performed, avoiding invalid operations.

[0093] More preferably, such as Figure 3 As shown, the step of converting the time-domain signal into electromyographic frequency domain features for each short-segment signal through Fourier transform or wavelet transform, and generating corresponding hash feature codes based on the electromyographic frequency domain features and a hash algorithm, includes:

[0094] S1021: Decompose the short-time signal segment into no fewer than 6 sub-bands using wavelet packet transform;

[0095] S1022: Calculate the corresponding energy entropy and band energy operator for each sub-band, wherein the band energy operator estimates the instantaneous energy of the signal by calculating the second-order difference of the signal;

[0096] S1023: Concatenate the energy entropy and frequency band energy operator of each sub-band to form a high-dimensional feature vector;

[0097] S1024: A set number of hash functions are generated using local sensitive hashing with random projection. The output of each hash function is calculated on the high-dimensional feature vector, and the set number of hash outputs are concatenated in order to form the corresponding hash feature code.

[0098] Compared with the traditional Fourier transform, the wavelet packet transform of this invention has a stronger ability to localize the time and frequency of non-stationary signals, and can retain time information while decomposing the frequency band, making it suitable for feature extraction of small signals.

[0099] When a subtle signal is submerged in noise, a single frequency band may not be able to highlight its characteristics. Through multi-sub-band decomposition, the energy of the subtle signal may be concentrated in a specific frequency band, while noise energy is dispersed across multiple bands. Sub-band feature analysis can improve the signal-to-noise ratio. For example, for weak postpartum pelvic floor muscle contractions, the energy changes in specific sub-bands may be more significant than the overall signal.

[0100] Energy entropy reflects the degree of disorder in the distribution of signal energy within a sub-band. Valid electromyographic signals typically have low energy entropy, while noise or interference has high energy entropy. For small signals, energy entropy can serve as a regularity indicator; even if the signal amplitude is low, if its energy is concentrated within a specific sub-band, it can still be identified as a valid signal.

[0101] The bandgap energy operator captures rapid energy changes by calculating the second-order difference of the signal, making it sensitive to the transient bursts of minute signals. For example, a slight contraction of the pelvic floor muscles may manifest as a short-term energy spike; the second-order difference amplifies this change, enhancing the detection capability of minute signals. Compared to traditional root mean square (RMS) calculations, the bandgap energy operator is more sensitive to signal changes and is suitable for detecting minute electromyographic activities with low amplitude but high rate of change.

[0102] The energy entropy and band energy operator of each sub-band are concatenated into a high-dimensional feature vector (e.g., 6 sub-bands → 12-dimensional vector), achieving a multi-dimensional representation of complexity-energy changes. For minute signals, this fusion can compensate for the shortcomings of a single feature. For example, a sub-band with low energy entropy but a small band energy operator value may indicate weak but organized electromyographic activity; if only energy entropy or the band energy operator as a single feature is used, such signals may be missed.

[0103] For minute signals, LSH can quickly compare their features with those of known valid signals. For example, the weak electromyographic signals of postpartum patients may differ in amplitude from standard signals, but their frequency domain feature distributions are similar; LSH can capture this similarity. The outputs of multiple hash functions are concatenated sequentially to form a hash feature code (such as a 64-bit binary string), where each hash bit represents the signature of the feature vector in a certain projection direction. This encoding method: preserves the local similarity of feature vectors, allows for rapid similarity calculation using Hamming distance; compresses the feature dimension (from high-dimensional to low-dimensional binary), reducing storage and computational costs, making it suitable for real-time processing.

[0104] More preferably, the step of performing similarity matching between the generated hash feature code and each hash feature in the hash feature library to determine the similarity between the hash feature code and each hash feature in the hash feature library includes:

[0105] Obtain the generated hash feature encoding, traverse all hash features in the hash feature library, and calculate the Hamming distance between them;

[0106] If the Hamming distance between the two is less than the first set value, then the corresponding hash feature is used as a candidate feature;

[0107] Construct a cost matrix between candidate features and each hash feature in the hash feature library, and dynamically program the filling of the cumulative cost matrix to obtain the dynamic time warped distance;

[0108] The similarity between the hash feature encoding and each hash feature in the hash feature library is determined based on the dynamic time rule distance.

[0109] In this invention, the hash features are encoded as binary strings (e.g., 64-bit / 128-bit), and the Hamming distance calculation is faster than the traditional Euclidean distance calculation. For large-scale hash feature libraries, the number of candidate features can be greatly reduced by using Hamming distance for fast filtering (e.g., Hamming distance < 10), thus reducing the amount of subsequent computation.

[0110] Locality Sensitive Hashing (LSH) ensures that hash codes of similar features have close Hamming distances. For minute signals, even if their amplitude or phase differs from the template, as long as the frequency domain feature distribution is similar, LSH can still map them to similar hash codes. Such candidate features can be preserved by relaxing the Hamming distance threshold. Minor electromyographic signals may have slight temporal shifts (such as different contraction initiation times), and traditional Euclidean distance cannot handle this temporal scaling. Dynamic time warping distance finds the optimal path through dynamic programming, allowing non-linear alignment on the time axis. For example, if the feature sequence of the signal to be matched is [A,B,C] and the template is [A,B,B,C], dynamic time warping distance can align the B of the signal to be matched with the two Bs of the template, calculating a more accurate similarity. For pelvic floor electromyographic signals, dynamic time warping distance can compensate for temporal variations caused by muscle fatigue or differences in nerve conduction velocity.

[0111] A cost matrix (e.g., 100×100) is constructed between candidate features and the template, where each element represents the difference between the two features at the corresponding time point. Dynamic programming is used to fill the cumulative cost matrix to find the globally minimum matching path, avoiding local optima. For subtle signals, this method captures their overall similarity to the template over time, rather than just local feature matching. For example, weak electromyographic signals from postpartum patients may be flatter in the time domain, but dynamic time warping distance can identify their similarity to the core frequency domain feature sequences of normal signals.

[0112] In pelvic floor muscle function assessment, this method can accurately identify subtle signals at different stages: Early postpartum rehabilitation: Patients' pelvic floor muscle contractions are weak, and the signal timing may be irregular. Dynamic time warping distance captures core features through non-linear alignment, avoiding misjudgments caused by timing deviations. Nerve injury repair monitoring: The subtle electromyographic signals generated during nerve regeneration may have timing differences from the standard pattern. Dynamic time warping distance can dynamically adapt to this change and assess the progress of repair.

[0113] More preferably, the intelligent recognition and detection method further includes:

[0114] When a valid signal is detected, the hash value of the corresponding valid signal is extracted and added to a temporary buffer. When the number of valid signals in the temporary buffer meets the requirements, the next step is executed.

[0115] Run the DBSCAN algorithm on all historical unclustered hash values ​​in the buffer, where the DBSCAN algorithm includes the neighborhood radius and the minimum number of samples;

[0116] If the Hamming distance between the hash of a valid signal and the centroid of a cluster in the feature library is not greater than the neighborhood radius, then the new signal is determined to be assigned to that cluster, and the cluster size and last update time are updated.

[0117] If the hash of the valid signal and the Hamming distance between the centroids of each cluster in the feature library are both greater than the neighborhood radius, and the number of neighborhood samples of the cluster containing the corresponding valid signal is greater than a set number, then a new cluster is created.

[0118] The solution in this invention sets a buffer capacity (e.g., 10-50 samples) and triggers clustering only when a sufficient amount of valid signal has accumulated. For minute signals, a single sample may be difficult to classify accurately due to its indistinct features, while a set of multiple samples can highlight its statistical characteristics. For example, the amplitude of the pelvic floor muscle micro-contraction signal in postpartum patients is only 20% of that of normal signals. A single sample is easily overwhelmed by noise, but the frequency domain energy distribution of five consecutive samples may show consistency (e.g., concentrated energy in the 30-80Hz range). This weak but recurring pattern can be identified through buffer accumulation.

[0119] The neighborhood radius set in this embodiment controls the neighborhood range of samples in the hash space. For small signals, appropriately increasing the radius (e.g., a Hamming distance of 15-20) can tolerate feature differences, clustering similar but weak signals into one class. For example, pelvic floor muscle micro-contraction signals from different patients may be scattered in the hash space due to individual differences, but after expanding the radius, they can be identified as the same pattern. Minimum number of samples: Sets the minimum number of samples required to form a cluster (e.g., MinPts = 5). For small signals, this parameter prevents noise points from being mis-clustered, ensuring that new clusters are created only when there is sufficient supporting evidence.

[0120] DBSCAN can discover clusters of arbitrary shapes, making it suitable for capturing complex distributions of minute signals. For example, the electromyographic signals of postpartum patients may have features scattered in the hash space due to muscle damage, which traditional spherical clustering methods such as K-means are difficult to identify effectively. However, DBSCAN can cluster these scattered but related signals into a class through density connectivity.

[0121] A new cluster is created when a new signal is more than a threshold away from the centroids of all clusters and there are enough samples in its neighborhood. This allows the system to discover subtle signal patterns that are not predefined. For example, during rehabilitation training, patients may develop new muscle recruitment strategies (manifested as unique weak electromyographic patterns) that cannot be covered by traditional feature libraries. Dynamic clustering can automatically identify and add such new patterns.

[0122] The solution in this invention combines a temporary buffer with DBSCAN, allowing the system to gradually learn and integrate newly emerging subtle signal patterns without retraining the entire model. Density clustering naturally resists outliers, and the buffer strategy further filters out occasional noise. The dynamically updated feature library can adapt to individual differences and signal changes during the recovery process, making it particularly suitable for long-term monitoring scenarios.

[0123] More preferably, the intelligent recognition and detection method further includes:

[0124] All valid signal segments are spliced ​​together according to time series and then input into the pelvic floor muscle resistance detection model for identification to identify whether there are abnormal segments.

[0125] The training data in the pelvic floor muscle resistance detection model includes positive sample data and negative sample data, wherein the negative sample data includes noise signals and non-target muscle contraction signals.

[0126] Positive samples in this invention include complete signal sequences of normal pelvic floor muscle contractions (such as the temporal pattern of standard contractions), used to learn the normal characteristics of physiological signals. Negative samples include noise signals: non-electromyographic signals such as electrode interference and motion artifacts; and non-target muscle contractions: contraction signals of non-pelvic floor muscles such as abdominal and gluteal muscles (such as unconscious abdominal straining by the patient during testing). Through training by comparing positive and negative samples, the model can distinguish between normal pelvic floor muscle activity sequences and abnormal signals (noise / non-target contractions), especially with more accurate judgment of abnormalities in minute signals—even if the signal is weak, if the temporal pattern conforms to normal physiological patterns, it is judged as valid; otherwise, it is judged as abnormal.

[0127] Adversarial detection models typically consist of a generator and a discriminator (e.g., a GAN architecture): the generator learns to produce sequences that closely resemble real pelvic floor muscle signals; the discriminator distinguishes between real and generated sequences and identifies anomalous segments. This adversarial training mechanism forces the model to capture deep features of the signal (such as the temporal dependence of subtle signals), avoiding reliance solely on surface features (such as amplitude). For example, noise may mimic subtle signals in a single segment, but lacks physiological temporal regularity at the sequence level, and will be identified as abnormal by the discriminator.

[0128] The solution in this invention can distinguish between minute signals and non-target muscle activity. Contraction signals of non-target muscles (such as abdominal muscles) may be similar to minute signals of pelvic floor muscles in single-segment features, but their temporal patterns differ. For example, pelvic floor muscle contraction sequences are often associated with changes in intra-abdominal pressure; abdominal muscle contractions, on the other hand, may exhibit independent high-frequency burst patterns. The adversarial model, by learning these sequence pattern differences, can reduce the misjudgment rate of minute contractions of non-target muscles.

[0129] Even if a valid signal segment is initially identified, short-term noise (such as momentary poor electrode contact) may still be present in the sequence. The adversarial model can identify anomalous segments that are inconsistent with the preceding and following signals through sequence consistency checks.

[0130] Since pelvic floor muscle dysfunction (such as relaxation or spasm) often manifests as temporal abnormalities in signal sequences (such as prolonged contraction cycles and residual electromyography during relaxation), adversarial models can learn the temporal patterns of normal sequences (such as contraction duration of 1-2 seconds and relaxation interval of 3-5 seconds) and provide early warnings for abnormal temporal sequences in subtle signal sequences (such as contraction lasting 5 seconds and no obvious relaxation period).

[0131] The solution of this invention sequentially splices together effective signal segments from a time series to reconstruct the complete process of pelvic floor muscle activity (such as a contraction-relaxation cycle). A single effective signal segment may only reflect local electromyographic activity (such as micro-contractions), while sequential splicing can capture the temporal correlation of the signals (such as gradual changes in contraction intensity and periodic patterns).

[0132] For minute signals, sequence splicing can integrate scattered weak activity into a recognizable pattern. For example, multiple minute contractions in a postpartum patient may present as a complete contraction cycle of "weak enhancement-peak-slow decay" in the sequence, while single-segment analysis would easily regard it as independent noise.

[0133] The adversarial model can identify whether the minute contraction-relaxation cycles in the pelvic floor muscle signal sequences of postpartum patients are complete, helping to assess the recovery of muscle neuromuscular control. For example, the contraction sequence of normal postpartum women should contain a clear initiation-enhancement-decrease process, while those with poor recovery may exhibit weak and disordered signal fragments. The pelvic floor muscle signal sequences of patients with stress urinary incontinence may show abnormal patterns of delayed contraction or weak contractions that cannot be maintained. The adversarial model can quantify these abnormalities through sequence analysis, providing objective evidence for clinical diagnosis.

[0134] This mechanism, through a process of sequence construction, adversarial learning, and anomaly recognition, enables the verification of the physiological temporal legitimacy of minute but effective signals. While improving signal purity, it provides a more accurate technical means for the objective assessment of pelvic floor muscle function.

[0135] The solution of this invention utilizes multi-channel and Xining electromyography (EMG) signal acquisition: a multi-channel surface electrode array is used to cover key areas of the pelvic floor muscles, simultaneously acquiring EMG signals from different locations. Through spatial filtering algorithms, the electrical activity of different muscle groups is separated, enabling precise localization and individual analysis of the distribution areas of fast-twitch and slow-twitch muscle fibers.

[0136] In practical implementation, dynamic thresholds can be automatically set based on the individual's resting electromyographic (EMG) signal characteristics. During subsequent contraction detection, effective EMG signals can be distinguished from noise in real time, improving data accuracy. For example, the threshold range can be dynamically adjusted based on the standard deviation of the resting signal to ensure the sensitivity and specificity of signal acquisition.

[0137] Furthermore, an assessment model combining the characteristics of fast-twitch and slow-twitch muscle fibers can be constructed. Parameters such as peak value of rapid contraction, stability of sustained contraction, and fatigue index of endurance contraction are used to generate a comprehensive score through a weighted fusion algorithm. This scoring system can intuitively reflect the functional status of fast and slow-twitch muscle fibers in the pelvic floor muscles, providing a basis for personalized rehabilitation programs.

[0138] In actual testing, the amplitude changes of pelvic floor electromyographic signals are small, which may reflect subtle coordination disorders of muscle fibers rather than obvious muscle weakness. This change has a significant impact during static maintenance (such as prolonged sitting or coughing), but traditional testing may focus more on large contractions. The method in this application greatly improves the likelihood of identifying the above conditions.

[0139] In practical implementation, the corresponding trend can be determined by identifying the aforementioned minute signals. For upward trends, the final state can be determined to assist in intelligent detection.

[0140] The intelligent identification and detection method for pelvic floor electromyography signals in this embodiment of the invention achieves comprehensive identification of pelvic floor electromyography signals by using a combination of short-time segmentation, frequency domain feature extraction, hash coding, and context model recognition. This greatly reduces the risk of effective signals being removed as noise and helps improve the accuracy of subsequent data identification.

[0141] Example 2

[0142] Please see Figure 4 , Figure 4 This is a schematic diagram of the intelligent recognition and detection system for pelvic floor electromyography signals disclosed in an embodiment of the present invention. Figure 4 As shown, the intelligent recognition and detection system for pelvic floor electromyography signals may include:

[0143] Receiver module 21: Used to receive the pelvic floor surface electromyography signals detected by the surface electromyography instrument, and to divide the continuously acquired pelvic floor surface electromyography signals into corresponding short-term signal segments through a sliding window;

[0144] Transformation module 22: used to convert the time domain signal into electromyographic frequency domain features for each short-time signal segment through Fourier transform or wavelet transform, and generate corresponding hash feature codes based on the electromyographic frequency domain features and hash algorithm;

[0145] Matching module 23: Used to perform similarity matching between the generated hash feature code and each hash feature in the hash feature library to determine the similarity between the hash feature code and each hash feature in the hash feature library; if the similarity between the two is within a first set interval, the corresponding short-term signal segment is determined to be a valid signal segment; if the similarity between the two is within a second set interval, the corresponding short-term signal is determined to be a signal segment to be confirmed; if the similarity between the two is within a third set interval, the corresponding short-term signal is determined to be an invalid signal.

[0146] Identification module 24: used to acquire surface electromyography (EMG) signals within a set time range before and after the signal segment to be confirmed, and input the acquired EMG signals into the signal recognition model for identification to determine whether the signal segment to be confirmed is a valid signal.

[0147] The intelligent identification and detection method for pelvic floor electromyography signals in this embodiment of the invention achieves comprehensive identification of pelvic floor electromyography signals by using a combination of short-time segmentation, frequency domain feature extraction, hash coding, and context model recognition. This greatly reduces the risk of effective signals being removed as noise and helps improve the accuracy of subsequent data identification.

[0148] Example 3

[0149] Please see Figure 5 , Figure 5This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of the present invention. The electronic device can be a computer, a server, etc. Of course, in certain cases, it can also be a mobile phone, tablet computer, monitoring terminal, or other smart device, as well as an image acquisition device with processing capabilities. Figure 5 As shown, the electronic device may include:

[0150] Memory 510 storing executable program code;

[0151] Processor 520 coupled to memory 510;

[0152] The processor 520 calls the executable program code stored in the memory 510 to execute some or all of the steps in the intelligent recognition and detection method of pelvic floor electromyography signals in Embodiment 1.

[0153] This invention discloses a computer-readable storage medium storing a computer program that causes a computer to perform some or all of the steps in the intelligent identification and detection method for pelvic floor electromyography signals in Embodiment 1.

[0154] This invention also discloses a computer program product, wherein when the computer program product is run on a computer, the computer performs some or all of the steps in the intelligent identification and detection method for pelvic floor electromyography signals in Embodiment 1.

[0155] This invention also discloses an application publishing platform, which is used to publish computer program products. When the computer program products are run on a computer, the computer executes some or all of the steps in the intelligent recognition and detection method for pelvic floor electromyography signals in Embodiment 1.

[0156] In various embodiments of the present invention, it should be understood that the sequence number of each process does not necessarily imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0157] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they can be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0158] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0159] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-accessible memory. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several requests to cause a computer device (which can be a personal computer, server, or network device, specifically a processor in the computer device) to execute some or all of the steps of the methods described in the various embodiments of the present invention.

[0160] In the embodiments provided by this invention, it should be understood that "B corresponding to A" means that B is associated with A, and B can be determined based on A. However, it should also be understood that determining B based on A does not mean determining B solely based on A; B can also be determined based on A and / or other information.

[0161] Those skilled in the art will understand that some or all of the steps in the various methods of the embodiments described can be implemented by a program instructing related hardware. This program can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0162] The above provides a detailed description of the intelligent identification and detection method, system, electronic device, and storage medium for pelvic floor electromyography signals disclosed in the embodiments of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for intelligent recognition and detection of pelvic floor muscle electrical signals, characterized in that, The method comprises the following steps: Receiving the pelvic surface electromyogram signals detected by the surface electromyogram instrument, and dividing the continuously collected pelvic surface electromyogram signals into corresponding short-period signal segments through a sliding window; Converting the time-domain signals into electromyogram frequency-domain features through Fourier transform or wavelet transform for each short-period signal segment, and generating corresponding hash feature codes according to the electromyogram frequency-domain features and a hash algorithm; the conversion of the time-domain signals into electromyogram frequency-domain features through Fourier transform or wavelet transform for each short-period signal segment, and the generation of corresponding hash feature codes according to the electromyogram frequency-domain features and a hash algorithm comprise the following steps: Decomposing the short-period signal segments into sub-bands with a number not less than 6 through wavelet packet transform; Calculating the energy entropy and the band energy operator of each sub-band, wherein the band energy operator is estimated by calculating the second-order difference of the signal; Splicing the energy entropy and the band energy operator of each sub-band to form a high-dimensional feature vector; Generating a set number of hash functions through locally sensitive hashing with random projection, calculating the output of each hash function for the high-dimensional feature vector, and splicing the set number of hash outputs in sequence to form a corresponding hash feature code; Matching the generated hash feature code with each hash feature in the hash feature library to determine the similarity between the hash feature code and each hash feature in the hash feature library; if the similarity between the two is in a first set interval, the corresponding short-period signal segment is determined as an effective signal segment, if the similarity between the two is in a second set interval, the corresponding short-period signal segment is determined as a to-be-confirmed signal segment, and if the similarity between the two is in a third set interval, the corresponding short-period signal segment is determined as an invalid signal segment; the matching of the generated hash feature code with each hash feature in the hash feature library to determine the similarity between the hash feature code and each hash feature in the hash feature library comprises the following steps: Obtaining the generated hash feature code, traversing all hash features in the hash feature library, and calculating the Hamming distance between the two; If the Hamming distance between the two is less than a first set value, the corresponding hash feature is taken as a candidate feature; Constructing a cost matrix between the candidate feature and each hash feature in the hash feature library, and filling a cumulative cost matrix through dynamic programming to obtain a dynamic time warping distance; Determining the similarity between the hash feature code and each hash feature in the hash feature library according to the dynamic time warping distance; Obtaining the surface electromyogram signals within a set time range before and after the to-be-confirmed signal segment, and inputting the obtained surface electromyogram signals into a signal recognition model for recognition to determine whether the to-be-confirmed signal segment is an effective signal; the obtaining of the surface electromyogram signals within a set time range before and after the to-be-confirmed signal segment, and the inputting of the obtained surface electromyogram signals into a signal recognition model for recognition to determine whether the to-be-confirmed signal segment is an effective signal comprise the following steps: Obtaining the surface electromyogram signals within a set time range before and after the to-be-confirmed signal segment; The time domain feature extraction module is used for time domain feature extraction of the surface electromyogram signal in a set time range to obtain corresponding time domain features, and the frequency domain feature extraction module is used for frequency domain feature extraction of the surface electromyogram signal in a set time range to obtain frequency domain features; The time domain features and the frequency domain features are flattened and spliced to form a joint feature vector, the hash feature code in the second set interval is converted into a dense vector through an embedding matrix, the joint feature vector and the dense vector are input into an attention layer, and an enhanced composite feature is obtained; The enhanced composite feature is input into a bidirectional LSTM network to obtain dynamic change information of the surface electromyogram signal in a time sequence, and an effective signal probability is output through a classification decision layer to determine whether the to-be-confirmed signal segment is an effective signal.

2. The method of claim 1, wherein the method further comprises: Before the continuous collected pelvic surface electromyogram signal is segmented into corresponding short-period signal segments through the sliding window, the method further includes: determining initial window parameters of the sliding window, the initial window parameters including a sampling frequency, an initial window length and an initial overlap rate; calculating sample entropy of a signal in a current window, if the current sample entropy is greater than a first set value, reducing the window length of the sliding window, if the current sample entropy is less than a second set value, increasing the window length of the sliding window, and if the current sample entropy is between the first set value and the second set value, keeping the initial window parameters; when the sample entropy changes, only when the sample entropy of continuous multiple sliding windows all changes in the same way, the window length adjustment is performed.

3. The method of claim 1, wherein the method further comprises: The intelligent recognition and detection method further includes: when an effective signal is detected, extracting a hash value of the corresponding effective signal, and adding the hash value of the effective signal to a temporary buffer area, and when the number in the temporary buffer area meets the requirement, performing the next step; running a DBSCAN algorithm on all historical unclustered hash values in the buffer area, wherein the DBSCAN algorithm includes a neighborhood radius and a minimum sample number; if the Hamming distance between the hash value of the effective signal and the centroid of a certain cluster in the feature library is not greater than the neighborhood radius, determining that the new signal belongs to the cluster, and updating the size and the last update time of the cluster; if the Hamming distance between the hash value of the effective signal and the centroid of each cluster in the feature library is greater than the neighborhood radius, and the neighborhood sample number of the cluster where the corresponding effective signal is located is greater than a set number, creating a new cluster.

4. The method of claim 1, wherein, The intelligent recognition and detection method further includes: splicing data of all effective signal segments in a time sequence, and inputting the data into a pelvic floor muscle antagonism detection model for recognition to identify whether there is an abnormal segment; training data in the pelvic floor muscle antagonism detection model includes positive sample data and negative sample data, and the negative sample data includes noise signals and non-target muscle contraction signals.

5. A system for intelligent recognition and detection of pelvic floor muscle electrical signals, characterized in that, The method includes: a receiving module configured to receive a pelvic surface electromyogram signal detected by a surface electromyograph, and segment the continuous collected pelvic surface electromyogram signal into corresponding short-period signal segments through a sliding window; The transformation module is configured to convert time domain signals into myoelectric frequency domain features for each short period signal segment by Fourier transform or wavelet transform, and generate corresponding hash feature codes according to the myoelectric frequency domain features and a hash algorithm. The short period signal segment is decomposed into sub-bands by wavelet packet transform, and the number of the sub-bands is not less than 6. The energy entropy and the band energy operator of each sub-band are calculated, and the band energy operator is estimated by calculating the second-order difference of the signal. The energy entropy and the band energy operator of each sub-band are spliced to form a high-dimensional feature vector. A set number of hash functions are generated by using a locally sensitive hash with random projection, the output of each hash function is calculated for the high-dimensional feature vector, and the set number of hash outputs are spliced in order to form a corresponding hash feature code. The matching module is configured to perform similarity matching between the generated hash feature code and each hash feature in the hash feature library to determine the similarity between the hash feature code and each hash feature in the hash feature library. If the similarity between the two is in a first set interval, the corresponding short period signal segment is determined to be an effective signal segment, if the similarity between the two is in a second set interval, the corresponding short period signal segment is determined to be a to-be-confirmed signal segment, and if the similarity between the two is in a third set interval, the corresponding short period signal segment is determined to be an invalid signal segment. The generated hash feature code is obtained, all hash features in the hash feature library are traversed, and the Hamming distance between the two is calculated. If the Hamming distance between the two is less than a first set value, the corresponding hash feature is taken as a candidate feature. A cost matrix between the candidate feature and each hash feature in the hash feature library is constructed, and a dynamic programming is filled to obtain a dynamic time warping distance. The similarity between the hash feature code and each hash feature in the hash feature library is determined according to the dynamic time warping distance. The recognition module is configured to obtain surface myoelectric signals in a set time range before and after the to-be-confirmed signal segment, and input the obtained surface myoelectric signals into a signal recognition model for recognition to determine whether the to-be-confirmed signal segment is an effective signal. The surface myoelectric signals in the set time range before and after the to-be-confirmed signal segment are obtained, and the obtained surface myoelectric signals are input into a signal recognition model for recognition to determine whether the to-be-confirmed signal segment is an effective signal. The time domain feature extraction module is used for time domain feature extraction on the surface electromyogram signal in a set time range to obtain corresponding time domain features, and the frequency domain feature extraction module is used for frequency domain feature extraction on the surface electromyogram signal in a set time range to obtain frequency domain features; The time domain features and the frequency domain features are flattened and spliced to form a joint feature vector, the hash feature codes in the second set interval are converted into dense vectors through an embedding matrix, the joint feature vector and the dense vectors are input into an attention layer, and an enhanced composite feature is obtained; The enhanced composite feature is input into a bidirectional LSTM network to obtain dynamic change information of the surface electromyogram signal in a time sequence, and an effective signal probability is output through a classification decision layer to determine whether the to-be-confirmed signal segment is an effective signal.

6. An electronic device, comprising: Comprise: a memory storing executable program codes; a processor coupled with the memory; the processor invokes the executable program codes stored in the memory to execute the intelligent identification and detection method of the pelvic floor muscle electromyogram signal according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, wherein the computer program enables the computer to execute the intelligent identification and detection method of the pelvic floor muscle electromyogram signal according to any one of claims 1 to 4.

Images