An intelligent prediction model training system for pelvic-abdominal coordination function state

Abstract

The present application relates to the technical field of computer neural network, and discloses a kind of intelligent prediction model training system of pelvic abdominal cooperative function state, comprising: time delay parameter calculation module, for determining discrete time delay index based on envelope cross-correlation;Asymmetric convolution coding network, index offset convolution layer is embedded to call index offset memory reading pointer;Collaborative feature fusion module, element-by-element multiplication is performed;Time sequence misplacement suppression module, annular shift negative sample is constructed and inhibitory penalty is applied, the present application realizes the physical cause alignment of heterogeneous signal under the premise of not increasing network depth by implanting deterministic time sequence bias in convolution operator bottom layer, and constructs cause decision boundary using counterfactual training strategy, eliminates false co-occurrence misjudgment and reduces computing overhead.

Description

Technical Field

[0001] This invention relates to an intelligent prediction model training system for pelvic and abdominal coordination function status, belonging to the field of computer neural network technology. Background Technology

[0002] In current computational tasks involving the synergistic relationship between abdominal pressure and pelvic floor electromyography (EMG) from multiple heterogeneous biological signals, existing computer systems typically employ dual-stream parallel feature extraction architectures such as convolutional neural networks (CNNs) or recurrent neural networks (RNNs). These model construction methods assume strong feature correlations among multi-channel data frames at the same sampling time, and design parallel convolutional kernel topologies to extract spatial features accordingly. However, when applying general synchronous computational architectures to electromechanical delay-based biosynergistic systems, there is a fundamental misalignment between the computational graph topology and the physical signal transmission patterns. From the establishment of the active source signal to the generation of action potentials by the response source signal, there is a physical transmission delay of tens to hundreds of milliseconds. Standard neural network synchronous convolutional mechanisms lack the ability to structurally represent asymmetric temporal causal relationships, forcing the feature fusion layer to search for correlations in non-causally aligned time slices. To compensate for these architectural defects, existing techniques typically significantly increase network depth or introduce global attention mechanisms to fit time-shift patterns, leading to a surge in model parameters and computational power consumption. Small sample data are highly susceptible to overfitting.

[0003] To address the issue of device integration, existing technologies have attempted to combine the two, but these are mostly limited to simple hardware stacking, lacking in-depth pathological collaborative control logic. For example, Chinese invention patent CN111297355A discloses a biofeedback device for pelvic floor muscle rehabilitation training. Although this device integrates abdominal and pelvic floor treatment modules, the technical concept still involves two independent control loops operating in parallel. During treatment, the intensity adjustment of abdominal electrical stimulation is based solely on a preset program or the patient's pain tolerance in an open-loop control, without establishing a real-time mapping relationship with the pelvic floor muscle state. Existing technologies face technical bottlenecks in processing inherently physically delayed temporal data: the general convolutional kernel topology lacks an endogenous adaptation mechanism for temporal causal bias, and the feature extractor cannot align heterogeneous signal segments with causal relationships at the physical level, resulting in feature aliasing and loss; the pure data-driven loss function optimization strategy lacks biomechanical logical boundary constraints, causing the model to misjudge non-specific random synchronous noise as collaborative features and lacking the ability to identify false co-occurrences; and the stacking of complex network structures to compensate for topological defects results in a huge waste of computing resources and storage space, making efficient deployment difficult in computing-constrained embedded environments.

[0004] Therefore, the technical problem to be solved by this invention is how to construct an intelligent prediction model training system for pelvic and abdominal coordinated functional states that directly adapts to the electromechanical delay characteristics of biological signals at the network topology level, ensures the causality of feature extraction, and reduces computational complexity. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A training system for an intelligent prediction model of pelvic and abdominal synergistic functional state, which runs on an electronic computing device, comprising:

[0006] The time delay parameter calculation module is used to acquire first time-series data and second time-series data that have a correlation, calculate the cross-correlation function of the low-frequency envelope of the first time-series data and the second time-series data within a preset time window, and determine the discrete time delay index based on the global maximum value of the cross-correlation function;

[0007] Asymmetric convolutional coding networks contain parallel active coding branches and passive coding branches;

[0008] The active encoding branch embeds an index-offset convolutional layer, which is used to call the discrete time-delay index as a fixed offset of the input feature map memory read address pointer when performing convolution operations, so that the calculation window of the convolution kernel is translated relative to the original time axis, generating an aligned first feature tensor; the passive encoding branch is used to process the second time-series data and generate the second feature tensor.

[0009] The collaborative feature fusion module is used to perform channel-level element-wise multiplication on the first feature tensor and the second feature tensor to generate a collaborative feature tensor that represents the degree of causal overlap.

[0010] The temporal misalignment suppression module is used to construct a negative sample sequence that circularly shifts the second temporal data along the time axis to a position outside the preset associated time limit. It calculates the feature activation values ​​of the asymmetric convolutional coding network for the negative sample pair composed of the first temporal data and the negative sample sequence, and uses the feature activation values ​​as an inhibitory penalty term to be superimposed on the loss function of backpropagation, so as to drive the network weights to perform numerical minimization in the non-associated temporal interval.

[0011] Preferably, the discrete time delay index is executed by the time delay parameter calculation module. The determination logic follows the following operational rules: ,in, The amplitude envelope of the first time series data. The amplitude envelope of the second time series data. The preset time delay search interval, For time indexing, This represents the sliding time.

[0012] Preferably, the logic for calling the memory read address pointer in the index-offset convolutional layer includes: during the convolution sliding window operation, the index of the input data is... Mapped to ,in For discrete time delay index, This is the sampling step size for the time axis; this operation is used to eliminate the time delay between the first time series data and the second time series data by changing the data index path without changing the values ​​of the convolution kernel weight matrix.

[0013] Preferably, the operation of the time-series misalignment suppression module in constructing the negative sample sequence includes: keeping the first time-series data unchanged, and indexing the time axis of the second time-series data. Mapped to ,in The total length of the sequence. The shift constant is greater than the preset association threshold, representing the modulo operation; this operation is used to generate constructed negative samples that destroy the original temporal association but retain the statistical distribution characteristics.

[0014] Preferably, the channel-level element-wise multiplication operation performed by the collaborative feature fusion module is specifically used to: perform point-to-point multiplication of each channel feature map of the first feature tensor with the feature map of the corresponding channel of the second feature tensor; this operation constructs a feature filtering mechanism so that when the high value region of the first feature tensor and the high value region of the second feature tensor coincide at the same time step, the corresponding collaborative feature tensor outputs a high response value.

[0015] Preferably, the feature activation value calculated by the temporal misalignment suppression module is specifically the L1 norm of the sum of the absolute values ​​of all elements in the collaborative feature tensor; the system is configured to minimize the L1 norm while minimizing the prediction error, thereby enabling the asymmetric convolutional coding network to suppress the output value of the feature fusion channel when processing negative sample pairs.

[0016] Preferably, the first time-series data is a multi-dimensional sensor data stream reflecting the excitation state of the system, and the second time-series data is a multi-dimensional bioelectric signal data stream reflecting the response state of the system. The time delay parameter calculation module is also used to perform low-pass filtering processing on the first time-series data and the second time-series data respectively before calculating the cross-correlation function, so as to extract the basic envelope data reflecting the changing trend.

[0017] Preferably, both the active and passive coding branches in the asymmetric convolutional coding network employ depthwise separable convolutional structures; wherein, the index-offset convolutional layer operates only on the channel-wise convolutional layer in the depthwise separable convolutional structure, and is used to complete temporal dimension index alignment while maintaining the independence of features between channels.

[0018] Preferably, the system also includes a confidence assessment module for monitoring the peak significance of the cross-correlation function; when the peak significance is lower than a preset threshold, the confidence assessment module generates a pause update instruction to temporarily stop the update of the discrete time-delay index and retain the index value of the previous frame.

[0019] Preferably, the asymmetric convolutional coding network is also connected to a fully connected regression layer, which is used to map the collaborative feature tensor to predicted values ​​representing the collaborative functional state of the system; the loss function consists of a weighted sum of a first sub-term and a second sub-term, where the first sub-term measures the difference between the predicted value and the labeled value, and the second sub-term is the L1 norm of the feature activation value.

[0020] Compared with the prior art, the beneficial effects of the present invention are:

[0021] 1. An asymmetric temporal bias structure is constructed in the feature extraction layer of the convolutional neural network. A preset or dynamically adjustable index offset is introduced between the active source encoding branch and the response source encoding branch to change the distribution of the receptive field of the convolutional kernel in the temporal dimension. This architectural improvement enables the topology of the neural network computation graph to match the physical law of electromechanical delay in the transmission of bioelectrical signals. By utilizing structured prior constraints, the timing of abdominal pressure occurrence and pelvic floor electromyographic response are physically aligned in the feature extraction stage. This avoids consuming a large number of network layers to fit nonlinear delay relationships. The collaborative feature fusion operation directly captures the energy of signals with causal correlations, avoiding the model forcibly fitting temporally misaligned multi-source data and generating false feature associations. This improves the interpretability and prediction accuracy of the model at the physiological mechanism level.

[0022] 2. A counterfactual contrast regularization mechanism is introduced during the model training phase to construct negative sample pairs that exceed the physical causal limit on the time axis. Combined with an inhibitory loss function, this drives the model to output a zero response when faced with signal combinations that do not conform to biomechanical temporal logic. The training strategy defines the physical causal boundary in the model decision space, enabling the neural network to identify cooperative features and identify statistical coincidences caused by suppressing random noise or artifacts. In long-term, high-noise actual monitoring data streams, this ensures that the model only activates signal segments that satisfy specific temporal causal logic, solving the problem of data-driven models misjudging non-specific synchronization signals as functional cooperative signals, and ensuring the specificity and stability of the model under complex operating conditions.

[0023] 3. An envelope cross-correlation-guided hyperparameter gating mechanism is adopted, using the calculation results of low-dimensional signal envelope cross-correlation as control instructions to dynamically adjust the read address pointers of neural network convolutional layers. The parameter control strategy directly injects the deterministic signal analysis results into the structural parameters of the deep learning model, constructing a coarse-grained guiding fine-grained adaptive calibration loop. Without increasing network depth and computationally complex attention modules, the model adapts to dynamic drift electromechanical delays under different subjects or physiological states. The preprocessing stage eliminates temporal uncertainty, and the backbone network focuses on high-dimensional feature morphological analysis, achieving large-span coverage of individual differences with limited computing resources and reducing the model deployment's dependence on hardware computing power. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall logical architecture and multi-dimensional data closed-loop flow of the system of the present invention;

[0025] Figure 2 This is a comparison chart of the timing alignment effects of heterogeneous signals with index offset in this invention;

[0026] Figure 3 This is a flowchart of the internal temporal processing interaction of the asymmetric convolutional coding network of the present invention. Detailed Implementation

[0027] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0028] This invention discloses an intelligent prediction model training system for pelvic and abdominal synergistic functional states, running on a computer device. It is used to construct and optimize a deep neural network model capable of characterizing biomechanical and electromechanical delay properties. The system mainly includes a heterogeneous temporal data normalization unit, an asymmetric convolutional coding network, a synergistic feature fusion module, and a temporal misalignment suppression module. The heterogeneous temporal data normalization unit acquires first and second temporal data of the subject during a preset action process through a data interface. The first temporal data is a sequence of abdominal pressure signals reflecting the active excitation state, and the second temporal data is a sequence of pelvic floor electromyographic signals reflecting the passive response state. Based on the differences in physical dimensions and dynamic range between the two types of signals, this unit uses a sliding window-based Z-Score normalization algorithm to preprocess the raw data stream. The system is set to a length of [missing information]. Calculate the mean of the signal amplitude within a time window. with standard deviation and the signal value at the current moment Convert to Standard scores in the form of, where To prevent division by zero errors, a minimal constant is used to generate a normalized two-channel tensor with uniform statistical distribution characteristics.

[0029] The asymmetric convolutional coding network, as a core component for feature extraction, includes parallel active and passive coding branches to process normalized first-time-series and second-time-series data, respectively. Addressing the inherent electromechanical delays in biomechanical signal transmission, the active coding branch embeds an index-offset convolutional layer. This layer maintains the weight matrix values ​​of the convolution kernel unchanged during convolution operations, while invoking a discrete-time index. Applying a fixed offset to the memory read address pointer of the input feature map causes a positive translation of the receptive field window of the convolutional kernel relative to the original input data on the time axis, i.e., at the computation time... When performing convolution output, the actual read time is... The input data, of which The discrete time delay index is the sampling step size. The corresponding physical time offset is set to match the normal electromechanical delay range of the human body, for example, within the range of 30 milliseconds to 100 milliseconds. This achieves physical causal alignment of heterogeneous signals at the underlying logic of feature extraction. Memory addressing hardening procedure: Define a circular buffer. and length Write pointer As the sampling clock monotonically increases, the read logic is embedded into AGU (Address Generation Unit) microinstructions, and the physical address calculation follows... ,in The base address of memory. The sampling step size, Given the memory storage unit width, software-level memcpy or tensor slicing operations are strictly prohibited. Zero-copy offsets are only achieved by modifying DMA (Direct Memory Access) descriptors or page table mappings, and latency parameters are specified. The cache line offset is directly mapped to the address bus, ensuring that timing alignment operations do not consume the computing power of the Vector Processing Unit (VPU). The system further includes a time delay parameter calculation module for calculating the aforementioned discrete time delay index. Before the asymmetric convolutional coding network is executed, this module performs low-pass filtering and absolute value rectification on the first and second time-series data with a cutoff frequency of 5 Hz to 10 Hz to extract the amplitude envelope that reflects the trend of signal energy change. and This module searches within a preset time delay interval. The cross-correlation function of two envelope sequences is calculated internally, and the sliding time corresponding to when the cross-correlation function reaches its global maximum is identified, which is then determined as the discrete time delay index. The data is then passed as a dynamic parameter to the index-offset convolutional layer to update the offset of the memory read pointer, ensuring that the convolution operation is locked at the temporal phase where the signal energy correlation is strongest.

[0030] When the collaborative feature fusion module is connected to the outputs of the active and passive coding branches, it generates a collaborative feature tensor representing the strength of pelvic-abdominal coordination. This module performs channel-level element-wise multiplication on the time-aligned first and second feature tensors to construct a feature filtering mechanism. When the excitation features represented by the first time-series data and the response features represented by the second time-series data both exhibit high activation at the same calibrated time step, the result of the multiplication operation outputs a high response value, thereby filtering out non-causal background noise and directly quantifying the effective strength representing the pelvic-abdominal coordination function. The computation graph instruction-level collapse logic sets a sparsity truncation threshold. (typical value) Hardware comparator array parallel scanning cooperative feature tensor Real-time generation of binary mask bitmap ,in Command issuing unit analysis When detected At that time, the weight loading instructions for the node in subsequent fully connected layers are intercepted, and NOP (no-operation) instructions are filled into the pipeline to replace the original multiply-accumulate (MAC) instructions. Simultaneously, the clock enable signal of the computation core is lowered (ClockGating), physically cutting off the forward propagation path of invalid features. The temporal misalignment suppression module is used to update network weights based on the backpropagation algorithm. Its core logic lies in constructing counterfactual negative sample pairs to suppress false co-occurrence. During the model training phase, this module keeps the first temporal data unchanged and indexes the time axis of the second temporal data. Mapped to ,in The total length of the sequence. For a shift constant greater than a preset correlation threshold, such as 500 milliseconds, this operation generates a negative sample sequence that destroys the original temporal causal correlation but retains the statistical distribution characteristics. The system inputs this negative sample sequence into an asymmetric convolutional coding network and calculates the L1 norm of the sum of the absolute values ​​of all elements in the generated counterfactual collaborative feature tensor. This L1 norm is superimposed on the total loss function as an inhibitory penalty term, driving the network to output feature values ​​close to zero when processing inputs that do not conform to biomechanical temporal logic, thereby establishing a decision boundary based on physical causality.

[0031] Example 1: In an industrial application scenario for screening postpartum pelvic and abdominal dysfunction, the data acquisition terminal simultaneously captures the first time-series data (abdominal pressure signal) and the second time-series data (pelvic floor electromyography signal) of the subject under multiple abdominal compression movements. Due to the objective limitations of the human neuromuscular conduction mechanism, there is an electromechanical delay of 30ms to 100ms between the establishment of abdominal pressure and the generation of antagonistic contraction response of the pelvic floor muscles. This causes the energy peaks of the two heterogeneous signals to exhibit asynchronous distribution on the uncalibrated physical time axis. Under this condition, the time delay parameter calculation module performs low-pass filtering and absolute value rectification processing on the raw data stream with a cutoff frequency of 5Hz to 10Hz to extract the amplitude envelope reflecting the basic trend. and And within the preset time delay search interval Internal cross-correlation function is calculated to lock the discrete time delay index corresponding to the global maximum. The index As the active encoding branch of the structured parameter injection asymmetric convolutional coding network, it drives the index-offset convolutional layer to shift the memory read address pointer of the input feature map backward when performing convolution operations. At each sampling point, this operation, while keeping the convolution kernel weight matrix unchanged, extends the receptive field of the convolution window to cover the physically causally aligned signal segment, thereby eliminating phase differences in the temporal dimension. The collaborative feature fusion module receives the first and second feature tensors processed by the above physical alignment and performs channel-level element-wise multiplication. Since the active excitation features and passive response features overlap in the calibrated time coordinate system, this multiplication operation transforms the originally dispersed single-channel signal intensity into a high-amplitude collaborative feature response. If the above index offset operation is not performed, the output of this multiplication operation will approach zero due to peak misalignment. To further eliminate false synchronization caused by environmental noise or motion artifacts, the temporal misalignment suppression module constructs counterfactual negative sample pairs during training, that is, keeping the first temporal data unchanged while indexing the time axis of the second temporal data. Mapped to , where the shift constant The system calculates the L1 norm of the feature activation values ​​generated by the network for the negative sample pair and adds it to the loss function to minimize the value. This keeps the model weights in an inhibited state when processing input data that does not conform to biomechanical temporal logic, thereby constructing a discriminative model that only responds to pelvic-abdominal synergistic behaviors with real physiological causal relationships.

[0032] Example 2: To verify the effectiveness and technical advantages of the intelligent prediction model training system for pelvic and abdominal coordinated functional status disclosed in this invention, multiple controlled trials were constructed. The experimental platform was deployed on a workstation equipped with an NVIDIA RTX 4090 GPU and an Intel Core i9-13900K CPU. The system environment was Ubuntu 20.04 LTS, implemented based on the PyTorch 2.0 framework. All experimental data came from a clinical subset of the publicly available dataset PhysioNet, which included simultaneous recordings of abdominal pressure IAP and pelvic floor electromyography (sEMG). A total of 5,000 action samples from 150 subjects were involved. During the data preprocessing stage, all signals were subjected to bandpass filtering from 5 Hz to 10 Hz to remove baseline drift and high-frequency noise.

[0033] This experiment designed four sets of comparative models to evaluate the contribution of different technical features to the final prediction performance. Control group A adopted a traditional dual-stream CNN architecture without any time delay processing mechanism, directly aligning the IAP and sEMG signals according to timestamps. Control group B introduced a global attention mechanism Transformer based on control group A, attempting to automatically capture long-distance temporal dependencies through self-attention weights. The sample group of this invention fully adopted the asymmetric time-delay convolutional coding network, time delay parameter calculation module, and temporal misalignment suppression module described in the aforementioned specific implementation. In addition, a partially missing control group C was set up. Although this group adopted an asymmetric convolutional structure, the counterfactual contrast regularization mechanism was removed, and only the conventional cross-entropy loss function was used for training. The experimental evaluation indicators included prediction accuracy ACC, F1 score, and noise resistance stability index SNR-Drop in low signal-to-noise ratio environment, which is the percentage decrease in performance after the input signal is superimposed with 10dB Gaussian white noise. Table 1 shows the performance comparison results of different models on the test set.

[0034] Table 1: Comparison of Performance Metrics for Different Model Architectures

[0035]

[0036] Referring to Table 1, control group A had the worst performance due to feature extraction misalignment caused by ignoring physiological electromechanical delays. While control group B improved performance through an attention mechanism, it incurred significant computational overhead and remained insufficiently accurate in capturing local time delays. Control group C improved accuracy and F1 score by introducing a physical alignment mechanism, validating the core value of asymmetric time-delay convolution. This invention's sample group, building upon control group C, further introduced counterfactual contrast regularization, achieving optimal performance not only on clean data but also demonstrating excellent noise resistance and stability, with an SNR-Drop of only 2.3%, far lower than other groups. To further verify the effectiveness of the time-delay parameter calculation module in handling individual differences... A sensitivity analysis experiment targeting time delay offset was designed, and a synthetic dataset containing different simulated electromechanical delays (ranging from 0ms to 200ms with a step size of 20ms) was constructed. The data showed that as the physical delay increased, the performance of the control group A showed a linear downward trend. When the delay exceeded 50ms, the accuracy dropped below 60%. In contrast, the accuracy of the sample group of this invention remained at a high plateau of over 92% throughout the entire physiological delay range of 0ms to 150ms. Only under non-physiological extreme delays exceeding 180ms did a slight decrease occur. This non-linear performance inflection point (180ms) is highly consistent with the upper limit of normal human electromechanical delay.

[0037] Example 3: This example combines Figures 1 to 3 This document describes a training system for an intelligent prediction model of pelvic-abdominal coordinated functional status, such as... Figure 1As shown, the system receives first time-series data (active excitation state data) and second time-series data (passive response state data). The data stream is transmitted to the time-delay parameter calculation module, which performs a low-frequency envelope cross-correlation operation to determine the discrete time-delay index, thereby generating the corresponding memory read pointer offset. Simultaneously, the first time-series data is input to the active encoding branch with an embedded index offset convolutional layer. By calling the time-delay index shift calculation window, an aligned first feature tensor is generated. The second time-series data is input to the passive encoding branch for conventional feature extraction, generating a second feature tensor. The collaborative feature fusion module receives the two tensors and performs channel-level element-wise multiplication to characterize the degree of causal overlap and finally outputs the intelligent prediction model of pelvic and abdominal synergistic functional state. In addition, the system includes a time-series misalignment suppression module, which receives the first and second time-series data, constructs a ring-shifted negative sample sequence, and calculates a suppression penalty term. This penalty term and the feature activation value output by the collaborative feature fusion module are input into the loss function optimization module. The model parameters are updated by minimizing the prediction error and the suppression penalty term.

[0038] like Figure 2 As shown, the horizontal axis represents time t in seconds, and the vertical axis represents the normalized signal amplitude. The solid line in the figure represents the first time-series data used as the baseline, i.e., the abdominal pressure signal. The dotted line represents the unprocessed raw pelvic floor electromyography (EMG) signal. It can be observed that the raw pelvic floor EMG signal exhibits a significant lag compared to the abdominal pressure signal. The dashed line in the figure represents the aligned pelvic floor EMG signal after an index offset operation. While maintaining the original morphological characteristics, the signal is shifted to the left along the time axis, ensuring that the peak position precisely coincides with the peak position of the abdominal pressure signal in phase. Figure 3 As shown, the processing flow begins at the input layer, which sends the first time-series data, namely abdominal pressure data, to the active encoding branch, and the second time-series data, namely pelvic floor electromyography data, to the passive encoding branch. The active encoding branch receives the data at a time delay index. Then, the input feature map is fed into the index-offset convolutional layer. Inside the index-offset convolutional layer, the memory read pointer offset and the convolution window time axis translation are adjusted sequentially to generate temporally aligned features and return them. The active encoding branch performs depthwise separable convolution processing on the aligned features and outputs the aligned first feature tensor. At the same time, the passive encoding branch performs standard convolution feature extraction and depthwise separable convolution processing on the second temporal data to generate the second feature tensor. Finally, the tensors generated by the two branches are converged to the feature output layer.

[0039] Example 4: Discrete time delay index in the time delay parameter calculation module To ensure the sufficiency of the logic, this embodiment provides a targeted explanation of the principle and transparency of the procedure for acquiring this parameter. In actual bioelectrical signal processing scenarios, directly calculating the cross-correlation function based on the original signal is often affected by high-frequency electromyographic noise and baseline drift, leading to inaccurate peak localization. Therefore, this embodiment constructs a constrained cross-correlation calculation model based on amplitude envelope for the first time-series data acquired, namely the abdominal pressure signal. Compared with the second time-series data, namely pelvic floor electromyography signals Preprocessing is performed using a fourth-order Butterworth low-pass filter with a cutoff frequency of 5Hz to 10Hz to remove high-frequency noise components from both signals. The absolute value of the result is then used to generate an amplitude envelope that reflects the trend of signal energy variation. and This preprocessing step eliminates the interference of fine waveform differences in the signal on phase determination, ensuring that the correlation calculation focuses on the overall movement pattern of muscle contraction.

[0040] Secondly, based on the amplitude envelope, the cross-correlation function is calculated, and a time delay search window is set to match the physiological electromechanical delay range. The value is typically taken as [20ms, 150ms]. Within this window, the normalized cross-correlation function is calculated. : ,in, To determine the discrete time step within the search window, an optimization algorithm is used to determine the optimal time-delay index, traversing all possible... Value, recognition When the global maximum value is reached And lock it as a discrete time delay index. The index value This method directly characterizes the time delay in which the signal energy correlation is strongest between the establishment of abdominal pressure and the electromyographic response during the action. This calculation process not only uses physiological prior knowledge to constrain the search space and avoid misjudgment in non-causal intervals, but also eliminates the influence of differences in absolute signal amplitude through normalization processing, ensuring the stability and consistency of parameter determination under different subjects and different signal intensities.

[0041] Example 5: Regarding the counterfactual comparison regularization module, in actual deployment, the parameters may be affected by negative samples. To mitigate the risk of training oscillations or convergence failure due to improper selection of this parameter, this embodiment provides supplementary explanations of the engineering calibration procedure for this parameter. Before deploying the system in a specific clinical environment, standardized boundary detection experiments must be performed to determine the optimal shift constant. The experiment set up a series of increasing... Candidate values, starting from 100ms, are incremented in 100ms increments up to 2000ms. For each candidate... The model was rapidly pre-trained for 50 epochs while keeping other hyperparameters constant, and the counterfactual sample suppression rate on the validation set was monitored, i.e., the proportion of negative samples correctly classified as non-cooperative states. Experimental data showed that when... When the time is less than 300ms, the suppression rate is less than 75% because the shifted signal still partially retains the physiological afterimage of the original action, making it difficult for the model to distinguish between true coordination and proximal artifacts; when After more than 500ms, the inhibition rate rapidly climbed to over 98% and entered a stable plateau region, at which point the system would... The default baseline value is set to 500ms, and adaptive adjustment logic is configured: if the descent gradient of the counterfactual sample loss value is detected to be lower than the preset threshold in the early stage of training, the system will automatically adjust the timeline. Increase by 100ms until the loss converges, thereby ensuring the effectiveness of the regularization constraint under different subject data distributions.

[0042] In addition to addressing the dynamic range drift issue that may occur in heterogeneous time-series data normalization units when faced with different batches of acquisition devices or individual differences among subjects, this embodiment supplements the on-site baseline calibration procedure. Before each model training or prediction task is initiated, the system performs a 30-second resting-state data acquisition process. During this period, the subject remains relaxed, and the system calculates the background noise baseline values ​​of abdominal pressure and pelvic floor electromyography signals in real time. In addition to the peak statistical distribution of the signal amplitude, the system uses this baseline data to construct a dynamic truncation threshold, selecting all subsequently acquired signals below this threshold. The portion of the fluctuation is considered invalid and set to zero, while the Z-score standardized mean is dynamically adjusted according to the peak statistical distribution. with standard deviation This procedure eliminates inconsistencies in signal amplitude dimensions caused by changes in electrode patch contact impedance or differences in subject body fat percentage, ensuring that the data input to the neural network is always statistically distributed within a standard normal distribution centered at zero with unit variance.

[0043] Example 6: Regarding the physical implementation logic of calculating the cross-correlation function based on the amplitude envelope in the time delay parameter calculation module, this example provides an engineering specification for the key digital signal processing parameters in this calculation process, including the calculation of the normalized cross-correlation function. To ensure optimal time delay resolution accuracy with limited computing resources, it is necessary to reasonably set the time delay search step size. With search window width Search step size The setting needs to balance computational complexity and time-delay resolution. According to the Nyquist sampling theorem, if the effective bandwidth upper limit of the original electromyographic signal is... Theoretically, the sampling interval should be less than Considering that electromyographic signal energy is mainly concentrated in the 20Hz to 150Hz frequency band, this embodiment will... Setting it to 1ms is much smaller than the shortest rise time of the physiological electromyographic response, sufficient to capture minute phase differences at the millisecond level. The search window width... Based on the physiological limits of the human body, the normal latency period from the increase in abdominal pressure to the reflexive contraction of the pelvic floor muscles ranges from 30ms to 100ms. To cover possible pathological prolongation or measurement errors, [the following is omitted as it is not explicitly stated in the original text]. Setting it to [0ms, 200ms] ensures that the search range includes all possible physiological causal intervals while excluding irrelevant long time delay interference.

[0044] Before performing cross-correlation calculations, to eliminate the interference of DC components and low-frequency drift on the correlation peak position, the amplitude envelope needs to be adjusted. and Perform mean removal processing, specifically by calculating the arithmetic mean of the signal envelope within the current analysis window. and Then, subtract the mean from the original envelope to obtain the centered envelope. and The cross-correlation function is calculated using the centered envelope, and its discretized calculation formula is modified as follows: ,in, To determine the total number of sampling points within the analysis window, For the corresponding discrete time delay index, through the above mean removal and normalization steps, the cross-correlation coefficient is... Limiting the time-delay index to the range [-1, 1] not only eliminates the influence of signal strength but also effectively suppresses spurious correlations caused by baseline fluctuations, thus enabling the time-delay index locked based on the global maximum value. It can accurately reflect the most essential dynamic coordinated phase relationship between two heterogeneous biological signals.

[0045] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0046] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A training system for an intelligent predictive model of pelvic and abdominal coordinated functional status, which runs on an electronic computing device, characterized in that, include: The time delay parameter calculation module is used to acquire first time-series data and second time-series data that have a correlation, calculate the cross-correlation function of the low-frequency envelope of the first time-series data and the second time-series data within a preset time window, and determine the discrete time delay index based on the global maximum value of the cross-correlation function; Asymmetric convolutional coding networks contain parallel active coding branches and passive coding branches; The active encoding branch embeds an index-offset convolutional layer, which is used to call the discrete time-delay index as a fixed offset of the input feature map memory read address pointer when performing convolution operations, so that the calculation window of the convolution kernel is translated relative to the original time axis, generating an aligned first feature tensor; the passive encoding branch is used to process the second time-series data and generate the second feature tensor. The collaborative feature fusion module is used to perform channel-level element-wise multiplication on the first feature tensor and the second feature tensor to generate a collaborative feature tensor that represents the degree of causal overlap. The temporal misalignment suppression module is used to construct a negative sample sequence that circularly shifts the second temporal data along the time axis to a position outside the preset associated time limit. It calculates the feature activation values ​​of the asymmetric convolutional coding network for the negative sample pair composed of the first temporal data and the negative sample sequence, and uses the feature activation values ​​as an inhibitory penalty term to be superimposed on the loss function of backpropagation, so as to drive the network weights to perform numerical minimization in the non-associated temporal interval.

2. The intelligent prediction model training system for pelvic and abdominal coordinated functional status according to claim 1, characterized in that, Discrete time delay indexing executed by the time delay parameter calculation module The determination logic follows the following operational rules: ,in, The amplitude envelope of the first time series data. The amplitude envelope of the second time series data. The preset time delay search interval, For time indexing, This represents the sliding time.

3. The intelligent prediction model training system for pelvic and abdominal coordinated functional status according to claim 1, characterized in that, The logic for calling memory access pointers in index-offset convolutional layers includes: during the convolution sliding window operation, the index of the input data is... Mapped to ,in For discrete time delay index, This is the sampling step size for the time axis; this operation is used to eliminate the time delay between the first time series data and the second time series data by changing the data index path without changing the values ​​of the convolution kernel weight matrix.

4. The intelligent prediction model training system for pelvic and abdominal coordinated functional status according to claim 1, characterized in that, The operation of the time-series misalignment suppression module in constructing the negative sample sequence includes: keeping the first time-series data unchanged, and indexing the time axis of the second time-series data. Mapped to ,in The total length of the sequence. The shift constant is greater than the preset association threshold, representing the modulo operation; this operation is used to generate constructed negative samples that destroy the original temporal association but retain the statistical distribution characteristics.

5. The intelligent prediction model training system for pelvic and abdominal coordinated functional status according to claim 1, characterized in that, The channel-level element-wise multiplication operation performed by the collaborative feature fusion module is specifically used to perform point-to-point multiplication of each channel feature map of the first feature tensor with the feature map of the corresponding channel of the second feature tensor. This operation constructs a feature filtering mechanism so that when the high value region of the first feature tensor coincides with the high value region of the second feature tensor at the same time step, the corresponding collaborative feature tensor outputs a high response value.

6. The intelligent prediction model training system for pelvic and abdominal coordinated functional status according to claim 1, characterized in that, The feature activation value calculated by the temporal misalignment suppression module is specifically the L1 norm of the sum of the absolute values ​​of all elements in the co-feature tensor; the system is configured to minimize the L1 norm while minimizing the prediction error, thereby enabling the asymmetric convolutional coding network to suppress the output value of the feature fusion channel when processing negative sample pairs.

7. The intelligent prediction model training system for pelvic and abdominal coordinated functional status according to claim 1, characterized in that, The first time-series data is a multi-dimensional sensor data stream reflecting the system's excitation state, and the second time-series data is a multi-dimensional bioelectric signal data stream reflecting the system's response state. The time delay parameter calculation module is also used to perform low-pass filtering processing on the first time-series data and the second time-series data respectively before calculating the cross-correlation function, so as to extract the basic envelope data reflecting the changing trend.

8. The intelligent prediction model training system for pelvic and abdominal coordinated functional status according to claim 1, characterized in that, Both the active and passive coding branches in the asymmetric convolutional coding network employ depthwise separable convolutional structures. The index-offset convolutional layer operates only on the channel-wise convolutional layer in the depthwise separable convolutional structure, and is used to complete temporal index alignment while maintaining the independence of features between channels.

9. The intelligent prediction model training system for pelvic and abdominal coordinated functional status according to claim 1, characterized in that, The system also includes a confidence assessment module, which monitors the peak significance of the cross-correlation function. When the peak significance is lower than a preset threshold, the confidence assessment module generates a pause update instruction to temporarily stop the update of the discrete time-delay index and retain the index value of the previous frame.

10. The intelligent prediction model training system for pelvic and abdominal coordinated functional status according to claim 1, characterized in that, The asymmetric convolutional coding network is also connected to a fully connected regression layer, which is used to map the cooperative feature tensor to the predicted value representing the cooperative functional state of the system; the loss function consists of a weighted sum of a first sub-term and a second sub-term, where the first sub-term measures the difference between the predicted value and the labeled value, and the second sub-term is the L1 norm of the feature activation value.

Images