Physiological state analysis method based on multi-sensor fusion

Abstract

The application relates to the technical field of physiological state analysis, in particular to a physiological state analysis method based on multi-sensor fusion. The application aims at the problem that the scale of standard data samples is limited and it is difficult to generalize across users in a specific scene. A static user portrait is used as a condition token to inject an identification network, and multi-strategy data enhancement is used to enable the model to maintain the adaptation ability to different user groups under the condition of small samples. A trend judgment module compares sessions based on a statistical baseline, and new users use the baseline of similar groups for cold start to avoid competing with the identification model for training resources. In the text generation link, structured data is used to lock the fact domain, and prompt words are used to define the expression forbidden zone. The output is double-checked through sensitive word filtering and rule mapping. Non-compliant expressions are automatically converted into safe general descriptions. A closed loop is formed in three aspects of small sample modeling, cross-user generalization and output compliance. Finally, a user physiological state record report is generated and output to the terminal.

Description

Technical Field

[0001] This application relates to the field of physiological state analysis technology, and in particular to a physiological state analysis method based on multi-sensor fusion. Background Technology

[0002] With the continuous development of personal health monitoring and interactive experience technologies, in recent years, there has been an increasing variety of smart terminal products for personal health management targeting women. Some of these products are positioned as assisting in the autonomous activation training of local muscle groups. These products typically focus on assisting users in autonomous training of local muscle groups as their core function. They use built-in sensors to sense mechanical and physiological signals during use and combine them with mobile terminal applications to provide real-time interactive guidance and progress summaries.

[0003] Due to the specific nature of data collection sites and usage scenarios, the number of users willing and able to complete a full session according to standardized procedures and authorize the use of their data is limited. Accumulating high-quality labeled samples is extremely difficult, and training a high-precision state recognition model from scratch faces a severe data bottleneck. Different users exhibit significant differences in local anatomical structures, neural control patterns, and voluntary contraction habits. Models trained under small sample conditions are prone to a sharp drop in recognition accuracy when facing new users, making cross-user generalization difficult to guarantee. Automatically generated text feedback poses compliance risks. If a general language model is used to directly derive user reports from data, the model may, due to its perceived illusion, output unauthorized content containing diagnostic, predictive, or treatment suggestions without factual support. This could easily trigger compliance risks in health-related scenarios involving private data.

[0004] Therefore, how to accurately identify the activation state of local muscle groups and generate compliant user feedback reports under the conditions of limited sample data and significant individual differences is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] The main objective of this application is to provide a physiological state analysis method based on multi-sensor fusion, which aims to solve the technical problem in the prior art of how to accurately identify the activation state of local muscle groups and generate compliant user feedback reports under the conditions of limited sample data and significant individual differences.

[0006] To achieve the above objectives, this application provides a physiological state analysis method based on multi-sensor fusion. The method includes the following steps: acquiring multimodal sensor data, including vaginal pressure data at multiple points, motion sensing data, vaginal pelvic floor muscle activity signals, temperature and humidity data, and vaginal pH data; inputting the multimodal sensor data into a state recognition model for window-level physiological state recognition, obtaining physiological state recognition results and derived indicators statistically generated based on the physiological state recognition results; comparing the derived indicators with the user's historical baseline data to obtain physiological state trend judgment results; constructing a structured context based on the physiological state recognition results, physiological state trend judgment results, and user profile; inputting the structured context into a large language model under preset prompt word constraints to obtain initial recorded text; performing filtering operations on the initial recorded text, including at least sensitive word filtering and compliance mapping verification, to obtain a user physiological state record report, which is then output to the user terminal.

[0007] Optionally, the above-mentioned input of multimodal sensor data into the state recognition model for window-level physiological state recognition, to obtain physiological state recognition results and derived indicators statistically generated based on the physiological state recognition results, includes: extracting spatiotemporal features from multi-point pressure data in the vagina after time axis alignment to obtain a pressure spatial feature sequence reflecting the activation intensity distribution of each sub-region of the vaginal pelvic floor muscle group; extracting temporal features from motion perception data and vaginal pelvic floor muscle group activity signals after time axis alignment to obtain a high-frequency temporal feature sequence; performing frame-level feature encoding from temperature and humidity data and vaginal pH data to obtain a low-frequency environmental feature sequence; performing cross-modal attention interaction based on the pressure spatial feature sequence, high-frequency temporal feature sequence, and low-frequency environmental feature sequence, and additively fusing it with the user profile token corresponding to the user profile to obtain a conditional fusion feature sequence; performing temporal modeling based on the conditional fusion feature sequence to obtain the physiological state probability distribution at each time step, wherein the user profile token participates in each layer of self-attention calculation; and performing time-weighted aggregation based on the physiological state probability distribution at each time step to obtain the physiological state recognition results and derived indicators.

[0008] Optionally, the above-mentioned time-weighted aggregation based on the physiological state probability distribution at each time step to obtain physiological state identification results and derived indicators includes: averaging the physiological state probability distribution at each time step along the time axis to obtain a session-level average probability distribution; taking the category corresponding to the highest probability in the session-level average probability distribution as the physiological state identification result; extracting statistical features based on the physiological state probability distribution at each time step, including at least one of the following: time proportion of each state category, number of state duration segments, state transition frequency, pressure balance at multiple points, variance or entropy of the probability distribution, trend indicators of environmental temperature and humidity changes, and trend indicators of vaginal pH changes; and generating derived indicators based on the statistical features.

[0009] Optionally, before acquiring the multimodal sensor data, the method further includes: acquiring a training sample set, which includes multiple sets of labeled multimodal sensor sample data, each set of sample data labeled with a window-level physiological state label; performing multi-strategy data augmentation on the training sample set to obtain an augmented training sample set, wherein the multi-strategy data augmentation includes at least one of random time offset, time-scale stretching and compression, sensor channel noise superposition, channel masking, and cross-user sample mixing; and constructing an initial model, which includes a multimodal feature encoding layer, a cross-modal fusion layer, and a Transformer backbone network, wherein the multimodal feature encoding layer contains a method for converting static data into static data. The user profile features are mapped to an embedding layer of user profile tokens; the sample data from the enhanced training sample set is input into the initial model to obtain the predicted physiological state probability distribution; based on the predicted physiological state probability distribution and window-level physiological state labels, the loss function value is calculated, and the parameters of the initial model are updated based on the loss function value until the training convergence condition is met to obtain the state recognition model. The loss function value includes a first loss term calculated based on the predicted probability distribution and labels of a single window, and a second loss term calculated based on the session-level predicted distribution obtained by aggregating the predicted probability distributions of all windows and the session-level state labels corresponding to the sample data.

[0010] Optionally, the above-mentioned comparison of derived indicators with user historical baseline data to obtain physiological state trend judgment results includes: calculating the deviation value based on the central trend baseline in the derived indicators and user historical baseline data, wherein the central trend baseline includes at least one of the following: peak pressure baseline at multiple points of vaginal pelvic floor muscles, average contraction frequency baseline, duration baseline of each physiological stage, environmental temperature and humidity baseline, and vaginal pH value baseline; determining the deviation level corresponding to the deviation value based on the dispersion baseline in the user historical baseline data; and generating physiological state trend judgment results based on the deviation value and deviation level, wherein the physiological state trend judgment results include a description of the direction of change, the magnitude of change, and the significance of change.

[0011] Optionally, the above-mentioned filtering operation on the initial recorded text, including at least sensitive word filtering and compliance mapping verification, to obtain a user physiological status record report includes: marking the initial recorded text based on a predefined sensitive word library to obtain the initial recorded text with sensitive word marking; performing replacement or deletion processing on the marked sensitive words based on the initial recorded text with sensitive word marking to obtain the filtered intermediate text; performing compliance mapping verification on the filtered intermediate text based on a predefined compliance mapping rule set to obtain non-compliant expression markers and corresponding non-compliant reasons in the filtered intermediate text; performing compliance expression correction on the filtered intermediate text according to the non-compliant expression markers and corresponding non-compliant reasons, replacing the non-compliant expressions with the corresponding allowed expressions in the compliance mapping rule set to obtain the corrected text with full field compliance; and generating a user physiological status record report based on the corrected text with full field compliance.

[0012] Furthermore, to achieve the above objectives, this application also proposes a physiological state analysis device based on multi-sensor fusion, the device comprising: The data acquisition module is used to acquire multimodal sensor data, which includes pressure data at multiple points inside the vagina, motion sensing data, vaginal pelvic floor muscle activity signals, temperature and humidity data, and vaginal pH data. The state recognition module is used to input multimodal sensor data into the state recognition model for window-level physiological state recognition, obtain physiological state recognition results, and derived indicators generated statistically based on the physiological state recognition results. The trend judgment module is used to compare derived indicators with the user's historical baseline data to obtain the physiological state trend judgment result; The prompt information generation module is used to construct a structured context based on the physiological state recognition results, physiological state trend judgment results, and user profile, and call a pre-trained general large language model to generate text under the constraints of preset prompt words to obtain the initial recorded text. The prompt message generation module is also used to perform filtering operations on the initial recorded text, including at least sensitive word filtering and compliance mapping verification, to obtain and output the user's physiological status record report.

[0013] Furthermore, to achieve the above objectives, this application also proposes a physiological state analysis device based on multi-sensor fusion. The physiological state analysis device based on multi-sensor fusion includes: a memory, a processor, and a physiological state analysis program based on multi-sensor fusion stored in the memory and executable on the processor. The physiological state analysis program based on multi-sensor fusion is configured to implement the steps of the physiological state analysis method based on multi-sensor fusion as described above.

[0014] In addition, to achieve the above objectives, this application also proposes a storage medium storing a physiological state analysis program based on multi-sensor fusion, which, when executed by a processor, implements the steps of the physiological state analysis method based on multi-sensor fusion as described above.

[0015] The proposed technical solutions (one or more) have at least the following technical effects: By injecting user profiles as conditional tokens into the recognition network and combining them with data augmentation, the model learns the differentiated feature expressions of different groups under small sample conditions, alleviating the generalization difficulties caused by sample scarcity and significant individual differences due to the privacy of the collection scenario. Through a trend judgment module, cross-session comparisons are performed using statistical baselines, and new users are cold-started using baselines of similar groups, achieving longitudinal trend analysis without competing for training resources, thus enhancing the continuity and personalization of the evaluation. In the text generation stage, structured data is used to lock the fact domain, prompt words are used to delineate prohibited areas, and sensitive word filtering and rule verification are implemented to suppress unauthorized statements caused by language model illusions, automatically downgrading non-compliant outputs to safe descriptions. These steps collaboratively constitute a processing chain from multimodal data collection to compliant report output, achieving accurate assessment and reliable feedback of local muscle group activation states under the constraints of limited samples and high individual differences. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating the physiological state analysis method based on multi-sensor fusion provided in this application embodiment. Figure 1 ; Figure 2 A flowchart illustrating the physiological state analysis method based on multi-sensor fusion provided in this application embodiment. Figure 2 ; Figure 3 A flowchart illustrating the physiological state analysis method based on multi-sensor fusion provided in this application embodiment. Figure 3 ; Figure 4 A schematic diagram of the structure of a physiological state analysis device based on multi-sensor fusion provided in an embodiment of this application; Figure 5This is a schematic diagram of the physiological state analysis device based on multi-sensor fusion provided in the embodiments of this application. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0020] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0021] As described in the background section, how to accurately identify the activation state of local muscle groups and generate compliant user feedback reports under the conditions of limited sample data and significant individual differences is a technical problem that urgently needs to be solved.

[0022] In view of this, this application provides a physiological state analysis method based on multi-sensor fusion. Addressing the issues of limited standard data sample size and difficulty in cross-user generalization in specific scenarios, it injects static user profiles as conditional tokens into the recognition network, combined with multi-strategy data augmentation, enabling the model to maintain adaptability to different user groups under small sample conditions. The trend judgment module uses statistical baselines for cross-session comparisons, and new users perform cold starts using baselines of similar groups, avoiding competition for training resources with the recognition model. In the text generation stage, structured data is used to lock the fact domain, and prompt words define expression restrictions. After output, the text undergoes dual verification through sensitive word filtering and rule mapping, automatically converting non-compliant expressions into safe-level general descriptions. This forms a closed loop at the three levels of small sample modeling, cross-user generalization, and output compliance, ultimately generating a user physiological state record report and outputting it to the terminal.

[0023] The execution subject of this application embodiment can be a physiological state analysis device based on multi-sensor fusion, or a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or a thermal management device capable of realizing the above functions based on a physiological state analysis device based on multi-sensor fusion. This embodiment does not specifically limit it in this way. The following uses a physiological state analysis device based on multi-sensor fusion as an example to describe this embodiment and the following embodiments.

[0024] Reference Figure 1 The physiological state analysis method based on multi-sensor fusion provided in this application includes steps S10 to S60: Step S10: Receive the raw multimodal sensor data and perform time axis alignment and adaptive noise reduction preprocessing on the raw multimodal sensor data to obtain standardized multimodal sensor data.

[0025] It should be noted that this solution is applied to a smart intimate health terminal product. The product combines a main body with a handheld component. Multiple pressure sensing units are arranged axially and circumferentially on the outer surface of the insertable part to collect contact pressure data from different sub-regions within the vagina. An integrated motion sensor captures the device's posture changes in three-dimensional space. Electrodes are also distributed on the surface to collect activity signals reflecting changes in the electrical activity or tissue impedance of the pelvic floor muscles. Furthermore, the product surface integrates temperature and humidity sensors and a pH sensor to collect data on the temperature, humidity, and pH levels within the vagina, respectively. The product communicates wirelessly with user terminals such as smartphones, providing real-time guidance and phased status feedback for muscle activation training via an application. The usage scenario is for self-directed pelvic floor muscle training and status assessment. In a private environment, users can follow the application's guidance to position the insertable part of the product at an appropriate depth and posture, and then execute various contraction, holding, and relaxation sequences prompted by the application, such as assessing endurance through continuous contraction and assessing coordination through rapid alternating contractions. The device collects the aforementioned multimodal physiological data simultaneously during the process, analyzes it, generates a user physiological status record report, and displays it on the terminal interface to help users understand the strength, control ability, and changing trends of their pelvic floor muscles.

[0026] It is understood that multimodal sensor data refers to multi-channel data reflecting the user's local muscle group activity and movement posture, synchronously collected by multiple sensors integrated into a working terminal device. This data includes at least multi-point pressure data within the vagina, motion sensing data, and pelvic floor muscle activity signals. Multi-point pressure data within the vagina is collected by multiple pressure sensing units arranged axially and / or circumferentially along the insertable portion of the device, reflecting pressure changes in different sub-regions within the vagina during muscle contraction or relaxation. Motion sensing data is collected by built-in motion sensors to record information such as the device's posture angle, acceleration, and angular velocity in three-dimensional space, reflecting the user's trunk movement state. Pelvic floor muscle activity signals are collected by electrodes distributed on the device surface, reflecting the activation level and rhythmic pattern of the pelvic floor muscles from dimensions such as electromyographic activity and impedance changes. Temperature and humidity data, and vaginal pH data are collected by corresponding environmental sensors at a lower sampling rate, reflecting the vaginal environment during user training. These data are synchronously collected by the device at a preset sampling rate and uploaded to the control terminal or cloud processing platform, forming the input basis for the entire analysis process of this application.

[0027] It should be noted that in this step, the raw multimodal sensor data uploaded by the device needs to undergo time axis alignment and adaptive noise reduction preprocessing before entering subsequent analysis to eliminate errors introduced by factors such as asynchronous acquisition by multiple sensors, hardware noise, and motion artifacts. This transforms the raw acquired signals into standardized data that can be directly used by the model. Because different sensors have different physical characteristics and sampling rates—for example, the sampling rate of pelvic floor muscle activity signals is usually higher than that of pressure signals, while the sampling rates of temperature, humidity, and pH data are much lower—if used directly without alignment, it will lead to time misalignment during subsequent cross-modal feature fusion, severely reducing recognition accuracy. Specifically, using the sampling clock of the vaginal pelvic floor muscle activity signal as the master clock, the pressure data, motion sensing data, temperature and humidity data, and vaginal pH data at multiple points within the vagina are unified to the same sampling rate and time reference through linear interpolation or resampling operations, resulting in a strictly time-aligned multi-channel synchronization matrix.

[0028] In terms of noise reduction, this solution employs differentiated adaptive noise reduction strategies tailored to the noise characteristics of different sensor signals, rather than using a single filtering method. For pressure data: First, a sliding median filter is used to remove spike pulses caused by momentary poor sensor contact. Then, piecewise least squares fitting is used to estimate and remove slowly varying baseline drift caused by sensor temperature drift or changes in the user's resting tension. The window length for piecewise fitting is adaptively determined based on the automatic detection results of the resting segment, ensuring that effective signals are not mistakenly eliminated during contraction activity. For pelvic floor muscle activity signals: A second-order Butterworth bandpass filter (passband 5-50Hz) is applied to retain the main spectral components of muscle activity and a 50 / 60Hz power frequency notch to eliminate power line interference. Simultaneously, the amplitude of acceleration output from the motion sensor is dynamically used to determine large-scale body movements. When the acceleration exceeds a preset threshold, an adaptive motion artifact suppression algorithm is activated. This algorithm utilizes the correlation between acceleration signals and electromyographic signals, subtracting motion-induced interference components from the electromyographic channel through an adaptive filter. For motion sensing data: a complementary filter is used to fuse accelerometer and gyroscope signals, leveraging the low-frequency stability of the accelerometer and the high-frequency response of the gyroscope to obtain smooth attitude angles and motion trajectories. For temperature and humidity data and vaginal pH data, due to their relatively slow changes, median filtering is used to remove occasional abnormal readings, followed by upsampling and alignment with the main time axis. Furthermore, a signal validity mask is generated during the preprocessing stage: electrode contact impedance and signal amplitude of each channel are monitored in real time. When the impedance exceeds the normal range or the signal exhibits clipping saturation, the corresponding time segment is marked as an invalid window. This mask is a binary sequence of the same length as the time-series data. The time segments marked as invalid are automatically excluded in subsequent state identification and derived index statistics, ensuring that the analysis results are not affected by poor sensor contact or extreme abnormal events. After the above preprocessing, standardized, low-noise, and synchronized multimodal sensor data is obtained, serving as input for subsequent window-level physiological state identification.

[0029] It should be understood that the collection, transmission, storage, and analysis of the aforementioned multimodal physiological data are all based on the user's prior explicit informed consent, and all data involving personal physiological information is protected using industry-standard encryption protocols during transmission and storage. Only when a user actively chooses to participate in the data sharing program through the application and makes a separate authorization can their de-identified data, which cannot be associated with a specific natural person's identity, be collected and used for training the model and improving the product involved in this application; the user has the right to withdraw the aforementioned authorization at any time, and withdrawal will not affect the legality of the processing activities carried out before withdrawal.

[0030] Step S20: Input the standardized multimodal sensor data into the state recognition model for window-level physiological state recognition to obtain physiological state recognition results and derived indicators generated based on the physiological state recognition results.

[0031] It should be noted that the state recognition model is a time-series classification model based on deep neural networks. It processes multimodal sensor data in units of data windows of preset time length, outputting the probability distribution of the physiological state category corresponding to that window. Window-level physiological state recognition refers to scanning the entire session's multimodal data stream with continuously sliding time windows, determining the user's pelvic floor muscle working state within each window, such as different state types like resting, mild contraction, maximum contraction, endurance contraction, or rapid alternating contraction. Derived metrics are a set of quantitative parameters calculated statistically based on the recognition results of all windows within a complete session. These parameters summarize the overall performance of the session, such as the time percentage of each state category, average contraction intensity, peak pressure level, contraction duration, and number of state transitions.

[0032] Understandably, introducing a window-level recognition mechanism can capture rapidly changing activation patterns of the pelvic floor muscles with fine-grained temporal resolution, avoiding the possibility of overlooking instantaneous state transitions in coarse-grained analysis based on entire sessions. During training and evaluation, users often experience various contraction types and rhythmic changes. Fine-grained recognition results not only provide continuous data for subsequent trend analysis but also enable the system to accurately determine whether the user has achieved sufficient duration or intensity in the target state, improving the accuracy of feedback.

[0033] It should be understood that the selection of the window size needs to strike a balance between resolution and stability. A window that is too short will introduce noise, while a window that is too long may blur the boundaries between different states. Typically, it can be set to several hundred milliseconds to several seconds, depending on the sampling rate and typical shrinkage period. At the same time, the calculation of derived indices should be performed after filtering out transient outlier state transitions in the identification results to improve the reliability of the statistics.

[0034] In one embodiment, before acquiring the multimodal sensor data, the method further includes: acquiring a training sample set, which includes multiple sets of labeled multimodal sensor sample data, each set of sample data being labeled with a window-level physiological state label; performing multi-strategy data augmentation on the training sample set to obtain an augmented training sample set, wherein the multi-strategy data augmentation includes at least one of random time offset, time-scale stretching and compression, sensor channel noise superposition, channel masking, and cross-user sample mixing; constructing an initial model, which includes a multimodal feature encoding layer, a cross-modal fusion layer, and a Transformer backbone network, wherein the multimodal feature encoding layer includes an embedding layer for mapping static user profile features to user profile tokens; inputting the sample data from the augmented training sample set into the initial model to obtain a predicted physiological state probability distribution; calculating a loss function value based on the predicted physiological state probability distribution and the window-level physiological state label, and updating the parameters of the initial model based on the loss function value until the training convergence condition is met to obtain a state recognition model.

[0035] It should be noted that the sensor channel noise superimposed in the above data augmentation strategy is used to simulate the weak random noise that may still remain after preprocessing, thereby improving the model's robustness to residual noise. This complements the noise reduction operation in the preprocessing stage. Preprocessing is responsible for eliminating large-amplitude, regular deterministic noise, while data augmentation allows the model to adapt to small-amplitude, random residual perturbations. The signal validity mask generated in the above preprocessing stage is also used during training: windows marked as invalid are not included in the loss calculation, preventing the model from learning incorrect feature-label mappings from unreliable data.

[0036] It should be noted that the above loss function values ​​include a first loss term and a second loss term. The first loss term is the classification cross-entropy loss between the predicted physiological state probability distribution of each window and the window-level physiological state label, weighted inversely according to the sample frequency of each category to alleviate the long-tail problem. The second loss term is the session-level distribution consistency loss, which is obtained by averaging the predicted probability distributions of all windows in the session to obtain the session-level predicted distribution, and calculating the KL divergence or mean squared error between it and the true global state distribution obtained by statistically analyzing the labels of all windows in the session. By introducing the second loss term, the model can ensure macro-control of the proportion of each state in the entire session while focusing on window-level local accuracy, further suppressing the model's overconfidence and misjudgment in dominant categories such as resting, and enabling the session-level average probability distribution to more accurately reflect the true state time allocation, providing a calibrated probability output for subsequent confidence assessment based on uncertainty indicators such as probability distribution entropy. The total loss is the weighted sum of the first and second loss terms, and the weight coefficients are determined through validation set tuning. The parameters of the initial model are updated based on the total loss value until the training convergence condition is met, resulting in the state recognition model.

[0037] It should be noted that the training sample set refers to a multimodal data set collected before model deployment, labeled with different contraction modes by multiple volunteers using equipment according to standardized procedures. Each sample corresponds to a data window and is accompanied by a label indicating the actual physiological state type of that window. Multi-strategy data augmentation aims to expand the diversity of training data under small sample conditions. Random time offset refers to randomly moving the window's starting point within a certain range; time-scale stretching and compression refers to linearly scaling the signal to simulate different contraction rates; sensor channel noise superposition refers to adding small-amplitude random noise to the original signal; channel masking refers to randomly obscuring parts of the sensor channel signal to simulate occasional sensor failures; and cross-user sample mixing refers to proportionally mixing window segments from different users to generate synthetic samples. User profile tokens are generated by mapping static user profile features (such as age range, reproductive history, training experience level, etc.) to a vector with the same dimension as the multimodal features through a dedicated embedding layer, which is then injected into the model as conditional information to participate in the attention calculation of each layer. The training convergence condition is typically that the loss function value no longer significantly decreases on the validation set or reaches the preset maximum number of iterations.

[0038] Understandably, user profiles are injected into the network backbone as "conditional tokens." This is key to solving the cross-user generalization problem. Static features such as age, height, weight, and ethnicity are no longer treated as ordinary input features, but are mapped to an independent user profile token. This token participates in the attention calculation of each layer of the Transformer backbone network. For example, this is equivalent to reminding the model at each stage, "The current user is facing a 40-year-old, slender user; please use a discrimination strategy suitable for this type of user." This strong prior condition significantly reduces the effective parameter search space of the model, making convergence and generalization possible under small sample conditions. Simultaneously, during the construction of the training sample set, all sample data have been authorized by users and de-identified to ensure that they do not contain information that can be associated with a specific natural person's identity. Through this training method, the model can effectively learn the differences in pelvic floor muscle activation patterns among different groups of people under conditions of limited sample size and high annotation costs. Furthermore, the data augmentation strategy significantly improves the model's robustness to acquisition noise, motion distortion, and individual sensor differences. User profile tokens are integrated across all layers of the model, providing a stable high-level semantic prior. This enables the model to make adaptive adjustments based on the user's demographic characteristics when inferring the state of a window, thereby alleviating the difficulty of cross-user generalization caused by differences in individual anatomy and neural control. This allows models trained with small samples to maintain high recognition accuracy for different users.

[0039] It should be understood that during training, the number of samples in each state category in the training set should be kept basically balanced to avoid class imbalance that could cause the model to shift towards the majority state. When mixing samples across users, the mixing ratio needs to be constrained to avoid generating physically impossible combinations of sensor data.

[0040] Step S30: Compare the derived indicators with the user's historical baseline data to obtain the physiological state trend judgment result.

[0041] It's important to note that user historical baseline data refers to a set of reference indicators pre-calculated based on the user's own or similar users' historical conversation records. This includes the central trend (such as mean and median) and dispersion (such as standard deviation and interquartile range) of each derived indicator. The central trend baseline can be the average level of derived indicators in the user's recent conversations, or it can be a statistical baseline extracted from group data of similar users when sufficient historical data for the user is unavailable. Specifically, this refers to the pelvic floor muscle-related baseline. Other baselines include, for example, environmental temperature and humidity baselines, which reflect the historical average levels of training environment temperature and humidity, and vaginal pH baselines, which reflect the historical average and range of vaginal acidity and alkalinity. The introduction of these environmental baselines allows trend judgment to identify indicator shifts caused by environmental changes or physiological cyclical fluctuations, avoiding misinterpretation as true changes in muscle group status. User historical baseline data is continuously updated using an exponentially weighted moving average, giving higher weight to recent sessions to accommodate long-term, gradual changes in user muscle strength. Simultaneously, physiological cycle stage information (such as menstrual cycles and postpartum recovery stages) included in the user profile is used to maintain independent baselines in stages, eliminating baseline shifts caused by cyclical fluctuations in hormone levels and reducing misjudgment rates. Comparing derived indicators from the current session with historical baselines yields the deviation of current performance from historical typical levels, i.e., the physiological state trend assessment result. The result typically includes the direction of change (improvement, stagnation, regression), the magnitude of change, and a description of significance.

[0042] Understandably, this step utilizes statistical comparison methods to achieve longitudinal trend analysis of different session results without increasing additional training data overhead. In the scenario of a new user's cold start, the system directly uses the baseline of the same group as an initial reference. As the user accumulates sufficient personal data, it gradually transitions to the individual baseline. Specifically, the weight of the same group baseline decreases linearly with the increase of the user's personal data volume. For example, when the number of personal historical sessions reaches a preset threshold, the baseline is completely dominated by personal data, achieving a smooth transition from group statistics to precise individual baselines. This mechanism allows the system to provide meaningful comparative feedback when the user first uses it, while avoiding competition between the trend analysis module and the recognition model for limited training resources, achieving lightweight and efficient personalized trend tracking.

[0043] It should be understood that, to ensure the reasonableness of trend judgments, user baseline data should be updated regularly, and systematic deviations caused by device firmware upgrades or sensor characteristic drift should be calibrated. Furthermore, for parameters in derived metrics that are susceptible to single, accidental factors, trend judgments should rely more on continuous changes over multiple sessions rather than on a single sharp increase or decrease.

[0044] In one embodiment, the above-mentioned comparison of derived indicators with user historical baseline data to obtain physiological state trend judgment results includes: calculating a deviation value based on the central trend baseline in the derived indicators and user historical baseline data; wherein, the central trend baseline includes at least one of the following: peak pressure baseline at multiple points of vaginal pelvic floor muscles, average contraction frequency baseline, duration baseline of each physiological stage, environmental temperature and humidity baseline, and vaginal pH value baseline; determining the deviation level corresponding to the deviation value based on the dispersion baseline in the user historical baseline data; and generating physiological state trend judgment results based on the deviation value and deviation level, wherein the physiological state trend judgment results include a description of the direction of change, the magnitude of change, and the significance of change.

[0045] It should be noted that the central trend baseline is established separately for different derived indicators. For example, the peak pressure baseline at multiple points of the vaginal pelvic floor muscles reflects the historical average level of peak pressure at each pressure sensing unit under maximum contraction; the average contraction frequency baseline reflects the historical average number of contraction events per unit time; and the duration baseline for each physiological phase reflects the typical level of the duration that can be maintained in each phase, such as endurance contraction or rapid contraction. The environmental temperature and humidity baseline reflects the historical average and normal fluctuation range of vaginal temperature and humidity during training; and the vaginal pH baseline reflects the historical average and normal fluctuation range of vaginal acidity and alkalinity. The dispersion baseline is used to measure the degree of fluctuation of the indicator's past data. The deviation value is the difference between the current derived indicator and the central trend baseline. The deviation level is a classification of the degree of deviation, for example, using standard deviation as the unit, divided into several levels such as normal, slight improvement / regression, significant improvement / regression, and substantial improvement / regression.

[0046] Understandably, by correlating deviation values ​​with the dispersion baseline, it's possible to statistically explain whether current changes exceed the historical normal fluctuation range, avoiding misinterpreting normal random fluctuations as meaningful trends. For example, if the peak pressure of a session is 0.5 standard deviations higher than the baseline mean, it indicates slight improvement; if it's 2 standard deviations higher, it indicates significant improvement. For environmental temperature, humidity, and pH values, this solution uses a two-sided deviation assessment: if the pH value deviates from the user's historical baseline by more than 2 standard deviations, the system marks it as "abnormal environmental fluctuation" rather than a change in muscle state. The resulting description of the direction and significance of the change provides users with a more scientific and reliable assessment of progress.

[0047] It should be understood that for new users who have just completed the cold start, their personal dispersion baseline has not yet been established. It can be temporarily replaced by the dispersion baseline of the same group of people, and gradually updated to personal dispersion after personal data has accumulated a certain number of conversations, so as to ensure the rationality of the deviation level classification.

[0048] Step S40: Construct a structured context based on the physiological state recognition results, physiological state trend judgment results, and user profile.

[0049] It's important to note that structured context refers to assembling the numerical and categorical analysis results output from previous steps according to a predefined data structure, forming machine-readable, easily processed information fragments for subsequent language model processing. This context typically includes: physiological state recognition results, indicating the main state distribution and derived indicator values ​​of the current session; physiological state trend judgment results, indicating the changes in each core indicator compared to historical data; and user profiles containing the user's basic attributes and training stage labels, such as age range, relevant training experience, and current training cycle. This information is organized in key-value pairs or similar formatted ways, limiting the fact domain upon which subsequent text generation is based and preventing the model from freely diverging from the raw data.

[0050] Understandably, the role of structured context is to strictly limit the input of subsequent large language models within a factual framework confirmed by system analysis, thus suppressing model illusions at the source. After transforming fuzzy, high-dimensional sensor data into clear, discrete analytical conclusions, the language model only needs to verbalize these conclusions without having to infer conclusions from complex data itself, significantly reducing the probability of generating biased or unauthorized statements.

[0051] It should be understood that the design of structured contexts needs to balance the completeness and conciseness of information. Important analytical dimensions should not be omitted, but overly detailed numerical listings should be avoided, which could lead to robotic or verbose text generation. During the construction process, the user's preferences can be used to determine which metrics are displayed.

[0052] Step S50: Under the constraints of preset prompt words, input the structured context into the large language model to obtain the initial recorded text.

[0053] It's important to note that cue word constraints are a set of pre-written and security-reviewed natural language instructions used to clearly inform the large language model of its role, expression boundaries, and content restrictions in this scenario. Cue words explicitly prohibit the model from outputting unauthorized content such as medical diagnoses, pathological judgments, treatment suggestions, or severity assessments that may cause anxiety. They also require the model to generate descriptive text based solely on facts provided within the structured context, without adding or supplementing unverified information. The initial recorded text is a user-oriented draft of training or evaluation records generated by the model based on the structured context, while adhering to the cue word constraints.

[0054] Understandably, cue word constraints, after locking in structured facts, define forbidden areas for expression, significantly reducing the likelihood of inappropriate content being generated, even in the face of the model's inherent tendency to produce illusions. For example, cue words explicitly require the model to describe trend changes as "improved since last time" or "compared to similar users," rather than diagnostic statements such as "returned to normal" or "suspected functional deficiencies." This constraint, together with the structured context, ensures that the model's output remains within a safe and controllable range at both the factual and expressive levels.

[0055] It should be understood that prompts should be reviewed and updated regularly to match evolving compliance requirements and product positioning. Furthermore, deployment should be adapted to different language versions to ensure semantic constraints are equivalent in cross-language scenarios.

[0056] Step S60: Perform filtering operations on the initial recorded text, including sensitive word filtering and compliance mapping verification, to obtain a user physiological state record report and output it to the user terminal.

[0057] It should be noted that sensitive word filtering refers to scanning the initial recorded text based on a predefined sensitive word library to identify and mark words or expressions that may violate privacy, ethical, or platform compliance requirements. Compliance mapping verification involves comparing expressions in the text against a set of compliance mapping rules. This rule set defines which expressions are non-compliant in a given context and specifies corresponding acceptable alternatives or safety fallback texts for each non-compliant expression. The user's physiological status record report is the final output generated after the above filtering operations. Its content maintains relatively natural readability while ensuring it does not contain any unauthorized or sensitive expressions. The user terminal can be a smartphone application interface, a device's built-in display, or a web page.

[0058] Understandably, these two layers of filtering constitute the third line of defense for ensuring output compliance. Sensitive word filtering directly blocks sensitive words at the word level, while compliance mapping verification identifies and corrects inappropriate expressions that may arise from oversight or implicit biases in the model at a deeper semantic level, such as describing increased muscle strength measured by the device as an implication of treatment effects. Through this series of processes, the output report provides valuable feedback and encouragement to users while always remaining within the boundaries of product safety and compliance.

[0059] It should be understood that the sensitive word database should be continuously updated to adapt to changes in the regulatory environment and public perception. The compliance mapping rule set can be jointly developed by legal or medical advisors and the development team. When the initial recorded text undergoes significant filtering and modification, a statement can be added at the end of the report stating "This report has undergone automatic security checks and adjustments in accordance with regulations" to enhance transparency.

[0060] This embodiment acquires multimodal sensor data and utilizes a state recognition model pre-trained with user profiles as conditional tokens and augmented with data to perform window-level physiological state recognition. The recognition results are compared with the user baseline to obtain trend judgments. Then, a structured context is constructed and initial record text is generated with the aid of prompt word constraints. Finally, after sensitive word filtering and compliance mapping verification, a safe and compliant user physiological state record report is output. This method organically integrates recognition, trend analysis, generation, and compliance filtering, achieving accurate state assessment and reliable feedback generation under conditions of small sample size and high individual variability.

[0061] In summary, this method addresses the challenges of sample scarcity and cross-user generalization through user profile token injection and data augmentation, achieves lightweight trend tracking through statistical baselines and cold start of the population, and suppresses model illusions and unauthorized expressions through three lines of defense: structured fact locking, cue word delimitation, and dual text filtering. Ultimately, it forms an end-to-end intelligent analysis solution suitable for private physiological scenarios, significantly improving the accuracy of state recognition and the security of output content.

[0062] Combination Figure 1 Please refer to Figure 2 In the physiological state analysis method based on multi-sensor fusion provided in this application embodiment, the above step S20 includes steps S201 to S205: Step S201: Based on the time-axis aligned and noise-reduced multi-point pressure data in the vagina, spatiotemporal features are extracted to obtain a pressure spatial feature sequence that reflects the distribution of activation intensity in each sub-region of the vaginal pelvic floor muscle group.

[0063] It should be noted that the intravaginal pressure data processed in this step has been time-synchronized and spatially calibrated. The data from each channel are strictly aligned at a uniform sampling rate, and baseline drift and spike noise have been filtered out. Spatiotemporal feature extraction uses spatial convolution to capture the pressure distribution of different channels at the same time, and models continuous changes in the time dimension through recurrent networks or temporal convolution, thereby obtaining a pressure spatial feature sequence that reflects the evolution of activation intensity in each sub-region over time. When the data from each pressure sensing unit are precisely aligned on the time axis, the pressure values ​​of different channels at the same time reflect the true mechanical distribution of the same contraction event in different sub-regions, rather than spurious spatial patterns caused by differences in sampling time.

[0064] It should be noted that the multi-point pressure data within the vagina is synchronously collected by multiple pressure sensing units arranged axially and circumferentially on the implantable outer surface of the aforementioned intelligent intimate health terminal product. Each pressure sensing unit corresponds to the contact pressure of different sub-regions within the vagina, collectively forming a multi-dimensional time series. Spatiotemporal feature extraction refers to processing these multi-channel pressure signals simultaneously in the temporal and spatial dimensions using convolution or attention operations. Specifically, at each time step, spatial convolution is first performed on the multi-channel pressure values ​​to capture the spatial topological relationship of the pressure in each sub-region at that moment. Then, along the time dimension, a recurrent network or temporal convolution is used to sequentially model the spatial features of multiple consecutive time steps, capturing the changing patterns of each sensing unit reading over time and the spatial interrelationships between pressure values ​​at different locations. The pressure spatial feature sequence is the extracted feature map sequence, with each frame encoding the spatial distribution information of the activation intensity of each sub-region of the pelvic floor muscle group (such as the anterior wall, posterior wall, left side, right side, etc.) at that time step.

[0065] Understandably, through spatiotemporal feature extraction, the model can distinguish between global synchronous contraction and local asymmetric activation patterns of the pelvic floor muscles. This is directly related to the muscle coordination performance of users when performing application-guided sequence training. Unlike methods that only perform independent temporal modeling of signals from each channel, spatiotemporal joint extraction can capture cross-channel coordination information such as "whether the right side follows synchronously when the pressure on the left side increases." For example, when the application guides users to perform continuous contractions to assess endurance, spatiotemporal features can reveal whether the user exerts force evenly throughout the body or only locally, thus laying the foundation for identifying contraction quality and providing personalized feedback. This spatial resolution capability allows subsequent state recognition to not only determine "whether a contraction is occurring," but also to further determine the symmetry and coordination of the contraction.

[0066] It should be understood that the number and layout of pressure sensing units directly affect the ability to distinguish spatial features. The product design has already reasonably covered the main sensing areas within the constraints of the placement area. Users should position the product at an appropriate depth and posture, and try to maintain a stable body posture to reduce spatial pressure distribution distortion caused by positional shifts or body movement. Zero-point calibration of the sensors should be performed before use to eliminate individual differences in resting tension baseline.

[0067] Step S202: Based on the motion sensing data aligned with the time axis and processed by noise reduction, and the activity signals of the vaginal pelvic floor muscles, perform temporal feature extraction to obtain a high-frequency temporal feature sequence.

[0068] It should be noted that motion sensing data is collected by motion sensors integrated within the product, reflecting changes in the device's posture angles, acceleration, and angular velocity in three-dimensional space, corresponding to the user's torso movement. Pelvic floor muscle activity signals are collected by electrodes distributed on the product surface, reflecting the activation level and rhythmic patterns of the pelvic floor muscles from dimensions such as electromyographic activity or tissue impedance changes. Both types of data have high sampling frequencies, capturing more refined muscle contraction dynamics than pressure signals. Temporal feature extraction utilizes methods such as one-dimensional convolution or recurrent neural networks to extract high-level features reflecting transient changes and rhythmic patterns from these high-frequency time series, forming a high-frequency temporal feature sequence synchronized with the pressure spatial feature sequence.

[0069] Understandably, motion-sensing data helps models distinguish whether pressure changes are caused by active contractions or passive compression resulting from overall body movement. Pelvic floor muscle activity signals provide direct information about muscle fiber electrical activity. Combining these two data points for temporal feature extraction can improve the accuracy of identifying complex patterns such as rapid alternating contractions and endurance contraction maintenance. For example, in app-guided rapid alternating contraction training, muscle activity signals clearly reflect the start and end times of each contraction, while motion-sensing data is used to eliminate artifacts introduced by unintentional posture adjustments by the user.

[0070] It should be understood that electromyography / impedancemetry electrode signals are easily affected by skin contact conditions and power frequency interference. Therefore, bandpass filtering and notch filtering are typically applied before feature extraction, and baseline calibration is performed to ensure that the extracted temporal features accurately reflect physiological changes. Users should position the insertable portion of the product in the correct posture to ensure good contact between the electrodes and the vaginal wall.

[0071] Step S203: Based on the temperature and humidity data and vaginal pH data, perform frame-level feature encoding to obtain a low-frequency environmental feature sequence.

[0072] It should be noted that the temperature and humidity data, as well as the vaginal pH data, were collected by environmental sensors integrated into the surface of the product's implantable portion. Compared to the aforementioned pressure, electromyography, and motion data, these two types of environmental parameters change at significantly slower rates. Within the timescale of a single training session, temperature, humidity, and pH typically exhibit only slow, monotonic changes or remain essentially constant. Therefore, this step employs a frame-level feedforward network encoding method. First, the environmental parameter sequence is upsampled to the same sampling rate as the main time axis. Then, it is mapped step-by-step through a fully connected layer to a fixed-dimensional feature vector, ultimately obtaining a low-frequency environmental feature sequence that is time-synchronized with the pressure spatial feature sequence and the high-frequency temporal feature sequence.

[0073] Understandably, the purpose of introducing environmental feature sequences is to provide environmental context information for the state recognition model, helping it distinguish between signal changes caused by muscle activity and sensor response shifts caused by changes in environmental conditions. For example, higher ambient temperatures may cause sensitivity drift in pressure sensing units or alter the user's basal metabolism, thus affecting the baseline amplitude of electromyographic signals. If the model lacks information on ambient temperature, it may misinterpret temperature-induced signal shifts as changes in the user's contraction intensity. Similarly, while fluctuations in vaginal pH are primarily driven by the menstrual cycle, extreme pH environments may affect the impedance characteristics of the electrode-tissue interface, indirectly interfering with the quality of electromyographic signal acquisition.

[0074] It should be understood that the encoding method for environmental features should match the physical characteristics of the signal. Temperature, humidity, and pH data change slowly, and their information density is much lower than that of high-frequency time-series signals. Lightweight frame-level feedforward networks can be used to fully extract their features without introducing complex time-series modeling, thus avoiding unnecessary computational overhead and the risk of overfitting.

[0075] Step S204: Based on the stress space feature sequence, high-frequency time series feature sequence and low-frequency environmental feature sequence, perform cross-modal attention interaction, and perform additive fusion with the user profile token corresponding to the user profile to obtain the conditional fusion feature sequence.

[0076] It should be noted that cross-modal attention interaction refers to inputting stress spatial feature sequences, high-frequency temporal feature sequences, and low-frequency environmental feature sequences into a cross-modal attention module. By calculating the attention weights among the three, cross-enhancement of spatial distribution information, dynamic rhythm information, and environmental context information is achieved. Specifically, stress spatial features are used as the query, and high-frequency temporal features and environmental features are used as the key and value, respectively, for cross-attention calculation. This allows the expression of stress spatial features to reference muscle activity patterns and environmental conditions at the same time. Simultaneously, high-frequency temporal features are used as the query, and stress spatial features and environmental features are used as key-value pairs for reverse cross-attention, allowing the expression of temporal features to incorporate spatial contraction distribution and environmental background. The user profile token is a representation obtained by mapping the user's static profile features (such as age range, training experience stage, etc.) to a fixed-dimensional vector through an embedding layer. Additive fusion adds the multimodal features after cross-modal interaction to the user profile token element-wise, directly injecting user-specific information into the expression of each layer of the network without changing the feature dimension and backbone structure.

[0077] Understandably, cross-modal attention interaction enables the model to answer the correlation between "where to contract" and "how to contract." For example, discovering a contradictory pattern of higher pressure on the left side and increased electromyography on the right side might indicate device bias or abnormal user control habits. Injecting user profile tokens essentially provides the model with a conditional prior, allowing it to adaptively adjust the judgment boundaries based on the typical characteristics of the user's group when determining the current contraction state. For example, for users with less training experience, the model might moderately relax the evaluation criteria for endurance contraction to improve the encouraging and usability of the feedback.

[0078] It should be understood that additive fusion requires the user profile token dimension to be strictly consistent with the multimodal feature dimension. During model design, the output dimension of the embedding layer must be equal to the feature dimension of the intermediate layers. The user profile token is optimized end-to-end as a set of learnable parameters during training and is pre-defined in the model parameters at the factory, requiring no user configuration.

[0079] Step S205: Perform time-series modeling based on the conditional fusion feature sequence to obtain the physiological state probability distribution at each time step, wherein the user profile token participates in the self-attention calculation of each layer.

[0080] It should be noted that the temporal modeling employs a Transformer backbone network with a self-attention mechanism at its core. Conditional fused feature sequences are input into this network sequentially. The self-attention computation at each layer allows direct interaction between features at any two time points, capturing long-range dependencies across windows in contraction events, such as the temporal progression of a complete contraction from exertion, reaching its peak, to gradual relaxation. The network's output at the last time step is a softmax-normalized probability distribution vector, representing the probability that the time window belongs to various predefined physiological state categories such as resting, mild contraction, maximum contraction, endurance contraction, and rapid alternating contraction. A user profile token is added as an additional special token to the sequence, participating in the self-attention computation at each layer, ensuring that feature aggregation during the temporal modeling process is globally influenced by the user's personalized prior knowledge.

[0081] Understandably, since a typical contraction of the pelvic floor muscles usually spans multiple time windows, direct temporal modeling can leverage the contextual information of preceding and following windows to improve the coherence and accuracy of current window state judgment, avoiding isolated state jumps. For example, a complete endurance contraction typically includes four phases: "exertion—holding—fatigue—relaxation." It's difficult to distinguish between "the slight decrease in strength due to fatigue in the later stages of holding" and "the beginning of voluntary relaxation" based solely on the pressure amplitude of a single window. Temporal modeling, however, can accurately determine which phase of contraction is currently in by utilizing the pressure change trends and electromyographic activity evolution patterns of preceding and following windows. User profile tokens are used throughout the entire temporal modeling process, enabling the model to recognize the differences among different user groups while integrating all historical window information. This helps achieve personalized recognition effects on top of a cross-user general model, improving the user's actual experience during training.

[0082] It should be understood that the computational cost of Transformer self-attention increases quadratically with the sequence length. When dealing with longer training sessions, inference latency can be controlled through engineering optimization techniques such as sliding window truncation, sparse attention, or linear attention, ensuring that analysis can be completed in real time or near real time on user terminals such as smartphones, without affecting the user's smooth experience due to model complexity.

[0083] Step S206: Perform time-weighted aggregation based on the probability distribution of physiological states at each time step to obtain physiological state identification results and derived indicators.

[0084] It should be noted that time-weighted aggregation refers to the statistical summarization of the probability distributions of all time windows within the entire training session to obtain a global conclusion at the session level. Typically, the probability distributions of each time step can be arithmetically averaged along the time axis to obtain the session-level average probability distribution, and the state category corresponding to the highest probability is taken as the primary physiological state recognition result for this session. Derived metrics, after obtaining the state sequence by hard decision based on the highest probability for each window, statistically calculate a series of quantitative indicators, including but not limited to: the time proportion of each state category (e.g., endurance contraction proportion, rapid contraction proportion), the number of state duration segments, the frequency of state transitions, the multi-point pressure balance index, environmental parameter statistics (e.g., session average temperature, average pH value), and uncertainty indicators such as the information entropy of the probability distribution.

[0085] Understandably, time-weighted aggregation transforms fine-grained frame-level probabilistic inferences into user-understandable summaries. The average probability distribution smooths out accidental misjudgments in individual windows, while multi-dimensional derived indicators provide rich quantitative characterization, directly applicable to subsequent trend analysis and report generation. For example, after guiding a user through an endurance assessment training session, the system statistically analyzes the percentage of time spent in endurance contraction and the longest duration of contraction—two derived indicators—to objectively measure the performance level of that endurance exercise. Furthermore, multi-point pressure balance reflects the symmetry of contraction, helping users perceive their own coordination weaknesses.

[0086] It should be understood that invalid windows before and after the formal start and end of training should be excluded when calculating derived indicators to avoid abnormally inflating the proportion of resting states. Multi-point pressure balance can be quantified by calculating the coefficient of variation of pressure values ​​in each channel or the pressure difference across each axis of symmetry. For sessions with high probability distribution entropy, it indicates a certain degree of uncertainty in model recognition; appropriate hints can be given in subsequent reports, such as "The signal was weak in some periods; the results are for reference only," to demonstrate transparency.

[0087] In one embodiment, the above-mentioned time-weighted aggregation based on the physiological state probability distribution of each time step to obtain physiological state identification results and derived indicators includes: averaging the physiological state probability distribution of each time step along the time axis to obtain a session-level average probability distribution; taking the category corresponding to the highest probability in the session-level average probability distribution as the physiological state identification result; extracting statistical features based on the physiological state probability distribution of each time step, including at least one of the following: time proportion of each state category, number of state duration segments, state transition frequency, pressure balance at multiple points, variance or entropy of the probability distribution, environmental temperature and humidity change trend indicators, and vaginal pH change trend indicators; and generating derived indicators based on the statistical features.

[0088] It should be noted that the session-level average probability distribution is a vector obtained by averaging the probability values ​​of each category across all windows. Each component reflects the overall probability of the corresponding state occurring throughout the session. The variance or entropy of the probability distribution reflects the degree of certainty of the model in each window. Small variance and low entropy indicate that the model is more confident in its state discrimination; conversely, large variance and entropy indicate that the signal is ambiguous and may contain mixed states or noise interference. Derived indicators are a set of indicators obtained by standardizing and quantifying these statistical features, such as "endurance contraction percentage," "average contraction symmetry index," and "state recognition confidence."

[0089] Understandably, using the average probability distribution instead of hard voting to obtain the main state can preserve the model's "hesitation" information about uncertain states, making the final judgment smoother. At the same time, the derived indicators incorporate uncertainty indicators such as entropy, which can provide users with prompts when generating reports later, such as "the signal was weak in some periods during this training, and the analysis results may have some bias", thereby increasing users' trust.

[0090] It should be understood that the probability distribution entropy and variance are significantly affected by the degree of model calibration. During model training, methods such as temperature coefficient calibration should be used to ensure that the output probability distribution accurately reflects the true confidence level. For low-frequency states, even if the average probability is low, their segment statistics may still be valuable for specific evaluations and should not be completely ignored.

[0091] In this embodiment, spatial features of pressure are extracted from multi-point pressure data within the vagina, and high-frequency temporal features are extracted from motion perception and pelvic floor muscle activity signals. Cross-modal attention interaction is introduced, and user profile tokens are additively fused. A Transformer network is used for temporal modeling, and finally, physiological state recognition results and derived indicators are obtained through time-weighted aggregation. Data augmentation and user profile token injection training strategies are employed during training. The entire process achieves a closed loop from multimodal perception to personalized state recognition.

[0092] In summary, this embodiment deeply integrates pressure spatial distribution, electromyographic rhythm, and user profile priors at the model structure level, enabling state recognition to possess spatial resolution and cross-user generalization ability. Even with limited sample conditions, it can still accurately distinguish the pelvic floor muscle activation patterns of different groups and provide highly informative derived indicators for subsequent trend analysis and report generation, effectively improving the personalization level of the product.

[0093] In combination with the above Figure 1 Please refer to Figure 3 The above step S60 includes steps S601 to S605: Step S601: Mark the initial record text based on the predefined sensitive word library to obtain the initial record text after sensitive word marking.

[0094] It should be noted that the predefined sensitive word library is a list of words established according to product compliance requirements and privacy protection regulations in the medical and health field. It includes words or phrases that may imply diagnostic conclusions, medical advice, sexual innuendo, emotional incitement, or excessive description of physiological parts. During the scanning and tagging process, the system performs word segmentation on the initial recorded text and matches it with the word library. It marks the location and type information of all matched words in the text, generating a text structure after sensitive word tagging, which is used for subsequent processing to locate and modify sensitive words.

[0095] Understandably, sensitive word filtering, as the first and fastest line of defense, can directly block obviously violating expressions from entering the final report at the word level, reducing the burden and risk of missed filtering in subsequent in-depth semantic verification. The use of the tagging method allows the system to selectively replace or delete downstream, rather than brute-force erasure, making it easier to achieve compliance correction while preserving text fluency.

[0096] It should be understood that the sensitive word database should support multilingualism and synonym variant coverage to prevent bypassing filtering through homophones, word splitting, or other methods. When tagging, attention should be paid to context-sensitive ambiguous words to avoid mistakenly penalizing expressions used reasonably in normal circumstances. The tagging results can be accompanied by contextual semantic tags for subsequent judgment steps.

[0097] Step S602: Based on the initial recorded text after sensitive word marking, perform replacement or deletion processing on the marked sensitive words to obtain the filtered intermediate text.

[0098] It should be noted that replacement processing refers to replacing the marked sensitive words with pre-specified neutral synonyms or broad descriptions, such as replacing potentially misleading intensity comparison words with "somewhat enhanced" or "remaining stable." Deletion processing refers to removing sensitive words and related phrases from the text when no suitable replacement can be found, and adjusting the context to make the sentences flow smoothly. The filtered intermediate text is the version of the text that has undergone compliance cleaning at the lexical level, and no longer contains obvious prohibited words, but may still hide non-compliance tendencies at the semantic level.

[0099] Understandably, replacement or deletion operations are minimal modifications made while maintaining linguistic consistency, avoiding a significant decline in report quality due to excessive filtering. For example, the original text might contain "visible improvement in pelvic floor muscle function," where "significant improvement in function" could be perceived as a therapeutic bias. Replacing it with "positive changes in pelvic floor muscle strength" conveys the information while avoiding compliance risks.

[0100] It should be understood that the replacement rules should be updated synchronously with the sensitive word database, and care should be taken to avoid creating new sensitive combinations or logical contradictions after replacement. For consecutive sensitive word segments, they can be processed together rather than replaced individually to maintain syntactic integrity.

[0101] Step S603: Based on the predefined compliance mapping rule set, perform compliance mapping verification on the filtered intermediate text to obtain the non-compliant expression markers and corresponding non-compliant reasons in the filtered intermediate text.

[0102] It should be noted that the compliance mapping rule set is a collection of conditional judgment rules. Each rule defines whether a certain expression pattern (such as a specific subject-verb-object combination, sentences with quantitative comparisons, inferential conjunctions, etc.) is compliant in a given context, and marks the category of non-compliance reasons (such as diagnostic bias, suggestive treatment, unauthorized personal health conclusions, etc.). During compliance mapping verification, the system performs syntactic analysis and semantic template matching on the filtered intermediate text. If any non-compliance rule is matched, the segment is marked as a non-compliant expression, and the corresponding non-compliance reason code is attached.

[0103] Understandably, this step delves into the semantic and grammatical levels, capturing inappropriate meanings or ambiguous risks remaining after simple sensitive word filtering, making it a crucial measure to ensure the security of output content. For example, the sentence "After this training session, your muscle strength has returned to normal levels" may not contain obvious sensitive words, but under compliance mapping rules, it would be judged as an implicit diagnostic conclusion, with the non-compliance reason marked as "returning to normal levels implies a pathological state." This semantic-level protection significantly improves the thoroughness of filtering.

[0104] It should be understood that the rule set needs to be continuously updated through the accumulation of real-world cases and human review feedback. Attention should be paid to the priority of rules to avoid conflicts caused by multiple rules triggering simultaneously. For low-confidence matching segments, automatic downgrading to human review or downgrading to more conservative, secure text can be implemented.

[0105] Step S604: Based on the non-compliant expression markers and the corresponding non-compliant reasons, correct the filtered intermediate text for compliant expressions, replacing the non-compliant expressions with the corresponding allowed expressions in the compliance mapping rule set, and obtain the corrected text that is compliant in all fields.

[0106] It should be noted that compliance statement correction refers to rewriting marked non-compliant segments into safe statements allowed by the rule using pre-defined replacement templates for each non-compliant rule. For example, statements with diagnostic bias are automatically corrected to descriptive statements, such as "pelvic floor muscle contraction strength reached 85% of the specified target" instead of "returned to normal". The corrected text with full field compliance is the final text form after dual processing of lexical filtering and semantic correction. All fields have passed compliance checks and can be directly used for report generation.

[0107] Understandably, automated compliance correction significantly reduces the workload of manual review while ensuring text consistency and timeliness. Since each correction is based on a pre-set security template, there are no secondary compliance risks. Even in the rare cases where a suitable template cannot be found, the system can call upon pre-set general security fallback text to ensure the output never goes astray. For example, it can automatically downgrade to "This training is complete; please see the chart for detailed data," ensuring users do not receive any potentially misleading content.

[0108] It should be understood that the revision process should follow the principle of minimal change, replacing only necessary sections and retaining most of the original wording to maintain the report's personalization and approachability to the greatest extent possible. The revised text can include a change log for easy tracking and review later.

[0109] Step S605: Generate a user physiological status record report based on the corrected text that complies with all fields.

[0110] It should be noted that the user's physiological status record report is the final product generated by combining and formatting the fully compliant, corrected text with necessary auxiliary elements (such as trend charts, key indicator values, training duration, etc.), and can be displayed to the user through the terminal interface. The report content includes both textual feedback and analytical summaries, as well as multi-dimensional quantitative indicators presented graphically, enabling users to understand the effectiveness of this training or evaluation from multiple perspectives.

[0111] Understandably, the report output in this step is within the final line of defense of the system's security boundary. Its language style, information granularity, and compliance level are all guaranteed, effectively supporting users' needs for continuous use and long-term self-management. Combining text analysis with visual charts reduces the burden on users reading text reports and allows descriptive sections that have been downgraded due to compliance processing to convey key information through charts. When generating the report, the modules displayed should be determined based on the user's historical preferences. For first-time users, a glossary can be included to improve user-friendliness. The report content should be stored encrypted and uploaded to the cloud only with user authorization to comply with privacy protection requirements.

[0112] In this embodiment, the initial recorded text generated by the large language model is processed sequentially, including lexical tagging and replacement / deletion based on a predefined sensitive word library, semantic verification and tagging based on a compliance mapping rule set, and compliance expression correction. Finally, a user physiological status record report with full field compliance is generated. The entire filtering process constructs two progressive lines of defense at the lexical and semantic levels. Combined with preset correction templates and fallback strategies, it achieves automated processing to output safe and reliable feedback text in private health scenarios.

[0113] In summary, the embodiments of this application, through the refined security design of the post-processing stage, enable intelligent private health terminal products to effectively suppress unauthorized expressions and compliance risks that may be caused by language model illusion when generating user reports based on multimodal data analysis, while maintaining text affinity and information integrity. This provides users with a thoughtful and safe status feedback experience during autonomous pelvic floor muscle training and assessment.

[0114] This application also provides a physiological state analysis device based on multi-sensor fusion, please refer to... Figure 4 The physiological state analysis device based on multi-sensor fusion includes: The data acquisition module 10 is used to acquire multimodal sensor data, which includes multi-point pressure data in the vagina, motion sensing data, vaginal pelvic floor muscle activity signals, temperature and humidity data, and vaginal pH data. The state recognition module 20 is used to input multimodal sensor data into the state recognition model for window-level physiological state recognition, obtain physiological state recognition results, and derived indicators generated based on the physiological state recognition results. The trend judgment module 30 is used to compare the derived indicators with the user's historical baseline data to obtain the physiological state trend judgment result. The prompt information generation module 40 is used to construct a structured context based on the physiological state recognition results, physiological state trend judgment results, and user profile; The prompt information generation module 40 is also used to call a pre-trained general large language model to generate text under the preset prompt word constraints, so as to obtain the initial recorded text; The prompt information generation module 40 is also used to perform filtering operations on the initial recorded text, including at least sensitive word filtering and compliance mapping verification, to obtain and output the user's physiological state record report.

[0115] The physiological state analysis device based on multi-sensor fusion provided in this application, employing the physiological state analysis method based on multi-sensor fusion in the above embodiments, can solve the technical problem of how to accurately identify the activation state of local muscle groups and generate compliant user feedback reports under conditions of limited sample data and significant individual differences. Compared with the prior art, the beneficial effects of the physiological state analysis device based on multi-sensor fusion provided in this application are the same as those of the physiological state analysis method based on multi-sensor fusion provided in the above embodiments, and other technical features in this physiological state analysis device based on multi-sensor fusion are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0116] This application provides a physiological state analysis device based on multi-sensor fusion. The physiological state analysis device based on multi-sensor fusion includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the physiological state analysis method based on multi-sensor fusion in the above embodiment 1.

[0117] The following is for reference. Figure 5 This document illustrates a structural schematic diagram of a physiological state analysis device based on multi-sensor fusion suitable for implementing embodiments of this application. The physiological state analysis device based on multi-sensor fusion in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 5 The physiological state analysis device based on multi-sensor fusion shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0118] like Figure 5As shown, the physiological state analysis device based on multi-sensor fusion may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 1002 or a program loaded from storage device 1003 into random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the physiological state analysis device based on multi-sensor fusion. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the multi-sensor fusion-based physiological state analysis device to wirelessly or wiredly communicate with other devices to exchange data. Although the figure shows a multi-sensor fusion-based physiological state analysis device with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0119] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0120] The physiological state analysis device based on multi-sensor fusion provided in this application, employing the physiological state analysis method based on multi-sensor fusion in the above embodiments, can solve the technical problem of how to accurately identify the activation state of local muscle groups and generate compliant user feedback reports under conditions of limited sample data and significant individual differences. Compared with the prior art, the beneficial effects of the physiological state analysis device based on multi-sensor fusion provided in this application are the same as those of the physiological state analysis method based on multi-sensor fusion provided in the above embodiments, and other technical features in this physiological state analysis device based on multi-sensor fusion are the same as those disclosed in the method of the previous embodiment, and will not be repeated here.

[0121] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0122] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0123] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the physiological state analysis method based on multi-sensor fusion in the above embodiments.

[0124] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0125] The aforementioned computer-readable storage medium may be included in a physiological state analysis device based on multi-sensor fusion; or it may exist independently and not assembled into a physiological state analysis device based on multi-sensor fusion.

[0126] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by a physiological state analysis device based on multi-sensor fusion, the physiological state analysis device based on multi-sensor fusion performs the following actions: acquires multimodal sensor data, including pressure data at multiple points within the vagina, motion sensing data, and vaginal pelvic floor muscle activity signals; inputs the multimodal sensor data into a state recognition model for window-level physiological state recognition, obtaining physiological state recognition results and derived indicators statistically generated based on the physiological state recognition results; compares the derived indicators with the user's historical baseline data to obtain a physiological state trend judgment result; constructs a structured context based on the physiological state recognition result, the physiological state trend judgment result, and the user profile; inputs the structured context into a large language model under preset prompt word constraints to obtain initial recorded text; performs filtering operations on the initial recorded text, including at least sensitive word filtering and compliance mapping verification, to obtain a user physiological state record report and outputs it to the user terminal.

[0127] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0128] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0129] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0130] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the aforementioned physiological state analysis method based on multi-sensor fusion. This solves the technical problem of accurately identifying the activation state of local muscle groups and generating compliant user feedback reports under conditions of limited sample data and significant individual differences. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the physiological state analysis method based on multi-sensor fusion provided in the above embodiments, and will not be repeated here.

[0131] The computer program product provided in this application can solve the technical problem of physiological state analysis based on multi-sensor fusion. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the physiological state analysis method based on multi-sensor fusion provided in the above embodiments, and will not be repeated here.

[0132] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A physiological state analysis method based on multi-sensor fusion, characterized in that, The physiological state analysis method based on multi-sensor fusion includes: Acquire multimodal sensor data, which includes pressure data at multiple points within the vagina, motion sensing data, vaginal pelvic floor muscle activity signals, temperature and humidity data, and vaginal pH data. The multimodal sensor data is input into the state recognition model for window-level physiological state recognition to obtain physiological state recognition results and derived indicators statistically generated based on the physiological state recognition results. The derived indicators are compared with the user's historical baseline data to obtain the physiological state trend judgment result; Based on the physiological state recognition results, the physiological state trend judgment results, and the user profile, a structured context is constructed; Under the constraints of preset prompt words, the structured context is input into the large language model to obtain the initial recorded text; The initial recorded text is subjected to filtering operations, including sensitive word filtering and compliance mapping verification, to obtain a user physiological state record report, which is then output to the user terminal.

2. The method according to claim 1, characterized in that, The step of inputting the multimodal sensor data into a state recognition model for window-level physiological state recognition, obtaining physiological state recognition results, and derived indicators statistically generated based on the physiological state recognition results includes: Spatiotemporal features were extracted from the time-axis aligned multi-point pressure data in the vagina to obtain a pressure spatial feature sequence that reflects the distribution of activation intensity in each sub-region of the vaginal pelvic floor muscle group. Based on the motion sensing data aligned to the time axis and the activity signals of the vaginal pelvic floor muscles, temporal features were extracted to obtain a high-frequency temporal feature sequence. Based on the temperature and humidity data and vaginal pH data, frame-level feature encoding is performed to obtain a low-frequency environmental feature sequence; Based on the stress space feature sequence, high-frequency time series feature sequence, and low-frequency environmental feature sequence, cross-modal attention interaction is performed, and additive fusion is performed with the user profile token corresponding to the user profile to obtain a conditional fusion feature sequence. Temporal modeling is performed based on the conditional fusion feature sequence to obtain the physiological state probability distribution at each time step, wherein the user profile token participates in the self-attention calculation of each layer. The physiological state identification result and the derived index are obtained by performing time-weighted aggregation based on the probability distribution of the physiological state at each time step.

3. The method according to claim 2, characterized in that, The step of performing time-weighted aggregation based on the probability distribution of physiological states at each time step to obtain the physiological state identification result and the derived index includes: The average probability distribution of physiological states at each time step is obtained by averaging along the time axis to obtain the conversation-level average probability distribution. The category corresponding to the highest probability in the session-level average probability distribution is taken as the physiological state recognition result; Based on the physiological state probability distribution of each time step, statistical features are extracted. The statistical features include at least one of the following: the time proportion of each state category, the number of state duration segments, the frequency of state transitions, the pressure balance of multiple points, the variance or entropy of the probability distribution, the trend index of environmental temperature and humidity changes, and the trend index of vaginal pH changes. The derived index is generated based on the statistical characteristics.

4. The method according to claim 1, characterized in that, Before acquiring multimodal sensor data, the process also includes: Obtain a training sample set, which includes multiple sets of labeled multimodal sensor sample data, each set of sample data labeled with a window-level physiological state label; The training sample set is subjected to multi-strategy data augmentation to obtain an augmented training sample set. The multi-strategy data augmentation includes at least one of random time offset, time-scale stretching and compression, sensor channel noise superposition, channel masking, and cross-user sample mixing. An initial model is constructed, which includes a multimodal feature encoding layer, a cross-modal fusion layer, and a Transformer backbone network. The multimodal feature encoding layer includes an embedding layer for mapping static user profile features to user profile tokens. The sample data from the enhanced training sample set is input into the initial model to obtain the predicted physiological state probability distribution; Based on the predicted physiological state probability distribution and the window-level physiological state label, calculate the loss function value, and update the parameters of the initial model based on the loss function value until the training convergence condition is met, thus obtaining the state recognition model. The loss function value includes a first loss term calculated based on the predicted probability distribution and labels of a single window, and a second loss term calculated based on the session-level predicted distribution obtained by aggregating the predicted probability distributions of all windows and the session-level state labels corresponding to the sample data.

5. The method according to claim 1, characterized in that, The step of comparing the derived indicators with the user's historical baseline data to obtain the physiological state trend judgment result includes: The deviation value is calculated based on the derived index and the central trend baseline in the user's historical baseline data. The central trend baseline includes at least one of the following: peak pressure baseline at multiple points of the vaginal pelvic floor muscles, average contraction frequency baseline, duration of each physiological stage baseline, environmental temperature and humidity baseline, and vaginal pH value baseline. Based on the deviation value and the dispersion baseline in the user's historical baseline data, the deviation level corresponding to the deviation value is determined; Based on the deviation value and the deviation level, the physiological state trend judgment result is generated, which includes the direction of change, the magnitude of change, and the significance of change.

6. The method according to claim 1, characterized in that, The step of performing filtering operations on the initial recorded text, including at least sensitive word filtering and compliance mapping verification, to obtain a user physiological state record report includes: The initial recorded text is marked based on a predefined sensitive word library to obtain the initial recorded text after sensitive word marking. Based on the initial recorded text after the sensitive words are marked, the marked sensitive words are replaced or deleted to obtain the filtered intermediate text; Based on a predefined set of compliance mapping rules, compliance mapping verification is performed on the filtered intermediate text to obtain non-compliant expression markers and corresponding non-compliance reasons in the filtered intermediate text. Based on the non-compliant expression markers and the corresponding non-compliant reasons, the filtered intermediate text is corrected to comply with regulations, replacing the non-compliant expressions with the allowed expressions corresponding to the compliance mapping rule set, to obtain the corrected text that is compliant in all fields; Based on the corrected text that complies with all fields, the user's physiological status record report is generated.

7. A physiological state analysis device based on multi-sensor fusion, characterized in that, The device includes: The data acquisition module is used to acquire multimodal sensor data, which includes multi-point pressure data in the vagina, motion sensing data, vaginal pelvic floor muscle activity signals, temperature and humidity data, and vaginal pH data. The state recognition module is used to input the multimodal sensor data into the state recognition model for window-level physiological state recognition, obtain physiological state recognition results, and derived indicators statistically generated based on the physiological state recognition results. The trend judgment module is used to compare the derived indicators with the user's historical baseline data to obtain the physiological state trend judgment result; The prompt information generation module is used to construct a structured context based on the physiological state recognition result, the physiological state trend judgment result, and the user profile; The prompt information generation module is also used to call a pre-trained general large language model to generate text under the preset prompt word constraints, so as to obtain the initial recorded text; The prompt information generation module is also used to perform filtering operations on the initial recorded text, including at least sensitive word filtering and compliance mapping verification, to obtain and output a user physiological state record report.

8. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the physiological state analysis method based on multi-sensor fusion as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, The computer program or instructions are executed by a processor to implement the physiological state analysis method based on multi-sensor fusion as described in any one of claims 1-6.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed, they implement the physiological state analysis method based on multi-sensor fusion as described in any one of claims 1-6.

Images