Model parameter updating method and system

Abstract

The application discloses a model parameter updating method and system, and a method implementation thereof, and comprises the following steps: when a preset physiological state detection model parameter updating trigger condition is met, acquiring physiological state detection results and physiological parameters of a target object in a target time period, wherein the trigger condition comprises user feedback triggering and model autonomous triggering; evaluating the physiological state detection results based on the physiological parameters to generate a self-evaluation adjustment coefficient; if the trigger condition is the user feedback triggering, obtaining a feedback confidence based on the matching degree between the physiological state detection results and user feedback information; and a model parameter updating unit is used to update the model parameters of the preset physiological state detection model according to the self-evaluation adjustment coefficient, the feedback confidence and a historical feedback information coefficient. The application can dynamically adjust parameters to adapt to complex scenes, optimize deviation reduction in multiple dimensions, reduce false negatives, false positives and missed detection, and improve monitoring accuracy and model robustness.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method and system for updating model parameters. Background Technology

[0002] In the field of sleep health monitoring, smart mattresses, smart base beds and other devices have widely adopted non-contact sensing technologies such as pressure sensors, bio-radar, and microphone arrays to collect physiological information such as snoring vibrations, respiratory rate, and body movement data. This eliminates the need for users to wear additional devices and better suits their daily sleep habits. However, most of these devices currently rely on preset fixed detection models to determine physiological states. Due to the complexity of sleep scenarios (such as differences in background noise in different bedroom environments, frequent changes in user sleep postures, and changes in bedding materials) and the lack of consistency in sensor devices (such as sensitivity deviations between different batches of sensors, performance degradation of sensors after long-term use, and differences in sensor fit during individual use), fixed models are difficult to adapt to the personalized monitoring needs of all scenarios and all users. This leads to frequent false detections (such as misjudging environmental noise as snoring, or identifying body movement interference as breathing abnormalities) and missed detections (such as missing low-intensity snoring or failing to capture brief breathing pauses) in certain special scenarios (such as monitoring weak snoring during light sleep or identifying breathing signals in noisy environments) or specific user groups (such as elderly people with abnormal pressure signals due to special body shapes or low respiratory rates). These issues seriously affect the accuracy and reliability of sleep monitoring results and cannot provide users with stable and accurate sleep health assessment data. Summary of the Invention

[0003] Therefore, it is necessary to provide a model parameter update method and system to address the above-mentioned technical problems and solve at least one of the problems existing in the prior art.

[0004] Firstly, a method for updating model parameters is provided, including:

[0005] When the preset triggering conditions for updating the parameters of the physiological state detection model are met, the physiological state detection results and physiological parameters of the target object within the target time period are obtained. The triggering conditions include user feedback triggering and model autonomous triggering.

[0006] The physiological state detection results are evaluated based on the physiological parameters to generate a self-evaluation adjustment coefficient;

[0007] If the triggering condition is user feedback trigger, a logical consistency and reliability assessment is performed based on the physiological state detection results, the physiological parameters, and the user feedback information.

[0008] The model parameters of the preset physiological state detection model are updated based on the self-assessment adjustment coefficient, feedback confidence level, and historical feedback information coefficient. The historical feedback information coefficient is determined by the trend and rate of change of the historical feedback detection deviation level.

[0009] In one possible implementation, after evaluating the physiological state detection results based on the physiological parameters and generating a self-evaluation adjustment coefficient, the method further includes:

[0010] If the triggering condition is model-initiated, the model parameters of the preset physiological state detection model are updated based on the self-evaluation adjustment coefficient and the historical feedback information coefficient.

[0011] In one possible implementation, the logical consistency and reliability assessment based on the physiological state detection results, the physiological parameters, and user feedback information includes:

[0012] Based on the physiological state detection results, the number of physiological state events detected and their corresponding probabilities for the target time period are obtained.

[0013] Based on the user feedback information, the number of actual physiological state events occurring within the target time period is estimated, and suspected physiological state data fragments are marked.

[0014] All suspected physiological state data segments are evaluated to obtain the number of suspected false physiological state segments reported by users;

[0015] The confidence level of the feedback is obtained based on the number of false positives for suspected physiological states reported by users and the number of times actual physiological state events occur.

[0016] In one possible implementation, updating the model parameters of the preset physiological state detection model based on the self-assessment adjustment coefficient, feedback confidence level, and historical feedback information coefficient includes:

[0017] The labels of suspected physiological state data segments are changed to physiological state labels, and the labels of suspected false physiological state data segments are changed to non-physiological state labels to form the first physiological state detection result.

[0018] Based on the first physiological state detection result, obtain the updated physiological state feature array and corresponding labels;

[0019] Obtain the historical physiological state feature array and corresponding labels, and combine it with the updated physiological state feature array and corresponding labels to update the dataset as parameters;

[0020] The dataset is updated based on the parameters, and the preset physiological state detection model is iteratively trained to obtain the updated preset physiological state detection model.

[0021] In one possible implementation, before updating the model parameters of the preset physiological state detection model based on the self-assessment adjustment coefficient, feedback confidence level, and historical feedback information coefficient, the following steps are included:

[0022] Based on the difference between the number of times the actual physiological state occurs and the number of physiological state events detected within the target time period, the feedback detection bias level is determined by a piecewise function.

[0023] The historical feedback detection deviation level is linearly fitted, and the slope of the fitted line is used as the trend of the historical feedback detection deviation level.

[0024] The rate of change is obtained based on the difference between the deviation level of the previous historical feedback detection and the deviation level of the current feedback detection;

[0025] The historical feedback detection deviation level change trend and change rate are weighted and fused, and then mapped through an exponential function to obtain the historical feedback information coefficient.

[0026] In one possible implementation, the step of evaluating the physiological state detection results based on the physiological parameters to generate a self-evaluation adjustment coefficient includes:

[0027] In the physiological state detection results, find all detection results that are determined to be physiological states and the duration of the physiological states;

[0028] Physiological state data segments whose duration is less than a preset time threshold are marked as suspected false physiological state segments.

[0029] Based on the physiological parameters, the remaining detection results that are determined to be physiological states are verified, and physiological state data segments that meet the preset false detection conditions are marked as suspected false detection physiological state segments.

[0030] Determine the number of all false positives labeled as suspected physiological states and the sum of the probabilities of all false positives labeled as suspected physiological states being identified as physiological states, and calculate the self-evaluation adjustment coefficient.

[0031] In one possible implementation, updating the model parameters of the preset physiological state detection model based on the self-assessment adjustment coefficient, feedback confidence level, and historical feedback information coefficient includes:

[0032] If the triggering condition is triggered autonomously by the model, then all labels marked as suspected false physiological states will be modified to non-physiological states, forming a second physiological state detection result;

[0033] Based on the second physiological state detection result, obtain the updated physiological state feature array and corresponding labels;

[0034] Obtain the historical physiological state feature array and corresponding labels, and combine it with the updated physiological state feature array and corresponding labels to update the dataset as parameters;

[0035] The preset physiological state detection model is iteratively trained based on the updated dataset of the parameters to obtain the updated preset physiological state detection model.

[0036] In one possible implementation, after updating the model parameters of the preset physiological state detection model based on the self-assessment adjustment coefficient, feedback confidence level, and historical feedback information coefficient, the method further includes:

[0037] The dataset is updated based on the parameters, and the updated preset physiological state detection model is evaluated to obtain the number of suspected false physiological state segments.

[0038] If the number of suspected false physiological state segments is less than a preset threshold, the historical physiological state feature dataset is updated.

[0039] If the number of suspected false physiological state segments is not less than the preset threshold, the model parameter update step size is adjusted.

[0040] If the number of times the model parameter update step size is adjusted is less than a preset threshold, the model is retrained based on the updated parameter dataset, and the number of suspected physiological state false detection segments in the updated model is evaluated.

[0041] If the number of times the model parameter update step size is adjusted is not less than the preset threshold, the update fails and the original model parameters are retained.

[0042] Secondly, a model parameter update system is provided, including:

[0043] On the device side, features are extracted from the acquired physiological data of the target object, and the extracted features are input into a preset physiological state detection model for prediction to obtain physiological state detection results.

[0044] The client is used to display the physiological state detection results to the user. If the user has feedback on the physiological state detection results, the user can input feedback information.

[0045] The server-side component, when a preset trigger condition for updating the parameters of the physiological state detection model is met, acquires the physiological state detection results and physiological parameters of the target object within a target time period. The trigger condition includes user feedback triggering and model self-triggered triggering. Based on the physiological parameters, the server evaluates the physiological state detection results and generates a self-evaluation adjustment coefficient. If the trigger condition is user feedback triggering, the server performs a logical consistency and reliability assessment based on the physiological state detection results, the physiological parameters, and user feedback information. The server updates the model parameters of the preset physiological state detection model according to the self-evaluation adjustment coefficient, feedback confidence level, and historical feedback information coefficient. The historical feedback information coefficient is determined by the trend and rate of change of historical feedback detection deviation levels.

[0046] In one possible implementation, the device includes:

[0047] The signal acquisition module is used to collect physiological data of the target object;

[0048] The signal quality judgment module is used to filter the collected physiological data to obtain target physiological data that meets the quality requirements.

[0049] The feature extraction module is used to extract features from the target physiological data to obtain physiological features;

[0050] The physiological state detection module is pre-configured with the preset physiological state detection model, which is used to predict the physiological characteristics and obtain the physiological state detection results.

[0051] The above-mentioned model parameter update method and system, the method implementation includes: when the triggering condition for updating the parameters of the preset physiological state detection model is met, acquiring the physiological state detection results and physiological parameters of the target object within a target time period, wherein the triggering condition includes user feedback triggering and model autonomous triggering; evaluating the physiological state detection results based on the physiological parameters to generate a self-evaluation adjustment coefficient; if the triggering condition is user feedback triggering, performing logical consistency and reliability evaluation based on the physiological state detection results, the physiological parameters, and user feedback information; updating the model parameters of the preset physiological state detection model according to the self-evaluation adjustment coefficient, feedback confidence, and historical feedback information coefficient, wherein the historical feedback information coefficient is determined by the trend and rate of change of historical feedback detection deviation level; and updating the model parameters of the preset physiological state detection model according to the self-evaluation adjustment coefficient, feedback confidence, and historical feedback information coefficient. In this embodiment, on the one hand, by combining user feedback triggering and model self-triggering for update, it can respond instantly when the user discovers obvious misjudgments, quickly correct sudden deviations, and actively capture scene changes or long-term data patterns. It can complete model iteration without manual intervention, which greatly improves the model's adaptability to complex sleep scenarios and individual user differences, and significantly reduces the probability of false detection and missed detection. On the other hand, by quantifying the fit between physiological parameters and detection results through self-evaluation adjustment coefficients, associating the matching between user feedback and detection results with feedback confidence, and fusing long-term deviation trends and short-term change rates with historical feedback information coefficients, the three factors work together to calculate the model parameter update step size and update value. This ensures that parameter adjustments not only fit the physiological logic of the current scenario but also retain historical effective monitoring experience, avoiding over-updates or under-updates. At the same time, the dynamic update mechanism can also adapt to the long-term fluctuations of user physiological characteristics (such as changes in snoring frequency) and sensor performance (such as sensitivity decay), enhancing model robustness and enabling the device to output stable and accurate detection results in all scenarios and throughout its entire life cycle. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a schematic diagram of an application environment for a model parameter update method in one embodiment of this application;

[0054] Figure 2 This is a flowchart illustrating a model parameter update method in one embodiment of this application. Figure 1 ;

[0055] Figure 3 This is a flowchart illustrating a model parameter update method in one embodiment of this application. Figure 2 ;

[0056] Figure 4 This is a schematic diagram of a model parameter update device in one embodiment of this application;

[0057] Figure 5 This is a schematic diagram of a model structure of a preset physiological state detection model in one embodiment of this application;

[0058] Figure 6 This is a schematic diagram of a computer device according to an embodiment of this application. Detailed Implementation

[0059] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0060] The model parameter update method provided in this embodiment can be applied to, for example, Figure 1 In the application environment of the model parameter update system, the system includes: device S1, client S2, and server S3, which can communicate and connect with each other. Device S1 can be a smart bed frame or smart mattress, and it mainly consists of four modules: signal acquisition module, feature extraction module, signal quality judgment module, and physiological state detection module. Client S2 includes, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices, which are pre-installed with physiological state detection-related apps or mini-programs. It may include a physiological state data display module and a physiological state detection feedback module. Server S3 can be implemented using a standalone server or a server cluster consisting of multiple servers. It may include a physiological state data storage module and a preset physiological state detection model parameter update module.

[0061] Specifically, device S1 collects physiological data of the target object through a signal acquisition module, filters valid data through a signal quality judgment module, extracts physiological features through a feature extraction module, and then predicts the valid data through a physiological state detection module to generate physiological state detection results. Device S1 then sends these results to server S3 and client S2. Server S3 stores the results through a physiological state data storage module, while client S2 presents the results to the user through a physiological state data display module. If the user has feedback on the results, they can submit feedback information in the physiological state detection feedback module of client S2, which then transmits the feedback information to server S3. Server S3 can trigger a model parameter update operation based on user feedback or automatic model updates through a preset physiological state detection model parameter update module. Finally, the updated model parameters are sent to device S1 so that device S1 can perform prediction operations using the latest model parameters. By working together to update model parameters, the system ensures that parameter adjustments not only align with the physiological logic of the current scenario but also retain historical and effective monitoring experience, avoiding over-updating or under-updating. At the same time, the dynamic update mechanism can adapt to long-term fluctuations in user physiological characteristics and sensor performance, enhancing model robustness.

[0062] Example 1

[0063] In one embodiment, such as Figure 2 As shown, a model parameter update method is provided, which is then applied to... Figure 1 Taking the server-side as an example, the explanation includes the following steps:

[0064] In step S110, when the preset triggering condition for updating the physiological state detection model parameters is met, the physiological state detection results and physiological parameters of the target object within the target time period are obtained. The triggering condition includes user feedback triggering and model autonomous triggering.

[0065] Physiological parameters refer to the raw / basic data that the device S1 directly collects through various sensors, which can reflect the physiological function or physiological activity characteristics of the target object (such as a sleeping user), such as audio signals, pressure signals, and respiratory signals.

[0066] The user feedback trigger refers to a passive update mechanism based on user-initiated intervention. When device S1 detects the physiological state of the target object, it sends the result along with the acquired physiological parameters to the client. The physiological state data display module presents the results to the user, allowing them to view the physiological state detection and assess for any discrepancies. These discrepancies can be obtained through feedback from a bed partner, audio recording analysis, or feedback from a reliable third-party device. If discrepancies exist, the user inputs feedback information through the physiological state detection feedback module, including any of the following: severe missed detection, missed detection, no missed detection, severe false detection, false detection, or no false detection. The feedback information, along with the corresponding physiological state characteristic data and physiological state detection results for the specified time period, is then uploaded to the server. The corresponding time period refers to the period from the previous day's fixed time (e.g., 6:00 PM) to the end of sleep. The physiological state detection results include whether snoring is present and its probability. It should be noted that snoring detection feedback can only be performed on the previous night's sleep.

[0067] The model-driven self-triggering mechanism refers to an active update mechanism that requires no user intervention. The system can have built-in algorithms to continuously monitor model metrics, such as detection result deviation indicators, data accumulation indicators, and update intervals. If the model's detection results for physiological states consistently fail to meet logical consistency with physiological parameters (e.g., "failing to meet standards for 3 consecutive detection cycles"), such as displaying the user's out-of-bed status but the model still determining there are physiological state detection results, then the model can be triggered to update autonomously. Alternatively, if the amount of new physiological parameters and detection result data within a target time period reaches a preset scale (e.g., the cumulative data volume reaches 100,000 records or one week's worth of sleep data), or if the time since the last parameter update exceeds a preset time threshold, such as 30 days, then the model can be triggered autonomously.

[0068] Optionally, based on the type of triggering condition, physiological state detection results and physiological parameters for a specified time period can be obtained. For example, if the current triggering condition is user feedback, the specified time period refers to a fixed time the previous day, such as 18:00, until the end of sleep, specifically determined according to the physiological state data cycle displayed on the client. If the triggering condition is model-triggered autonomously, the specified time period refers to a fixed time the previous day until a fixed time on the current day. Simultaneously, historical physiological state detection feedback data and historical physiological state detection results can also be obtained. Historical physiological state detection results may include feedback data from the most recent preset number of times, which can be 10. Historical physiological state detection feedback data may include the user's feedback status, differences in physiological state detection results within the feedback time period before and after the model update (the number of changes in detection results before and after the model update, the number of non-physiological states judged as physiological states, the number of physiological states judged as non-physiological states, and the ratio of the change in the number of detected physiological states to the number of detected physiological states before the model update).

[0069] In step S120, the physiological state detection results are evaluated based on the physiological parameters to generate a self-evaluation adjustment coefficient;

[0070] Optionally, key features related to the detection results are first extracted from physiological parameters (e.g., when snoring is detected, the periodicity of the audio signal and the synchronicity of the vibration signal are extracted). Then, the logical consistency and reliability of the features and detection results are compared according to physiological laws (e.g., if there is a conclusion of snoring but no corresponding periodic audio, the fit is poor). Finally, by statistically analyzing the proportion of mismatched segments and calculating the feature deviation, the fit is converted into a coefficient of 0-1 (or a limited range), thus obtaining the self-evaluation adjustment coefficient. It should be noted that the larger the coefficient, the worse the fit between the detection results and physiological parameters (lower credibility), and the more significant the subsequent model parameters need to be adjusted to correct the deviation; the smaller the coefficient, the better the fit (higher credibility), and the smaller the subsequent model parameter adjustment to ensure model stability. This coefficient is ultimately coordinated with the feedback confidence and historical feedback information coefficients.

[0071] The self-assessment adjustment coefficient is generated as follows: Based on the physiological state detection results, all detection results identified as the target physiological state (such as snoring) and their duration are identified. Data segments with a duration less than a preset threshold TH8 (e.g., 12 seconds) are marked as suspected false detections. Then, physiological parameters such as the bed-leaning signal from the pressure sensor and the activity status from the motion sensor are combined to determine the target object's state (e.g., bed-leaning, sleep state) in the remaining data segments, thereby identifying all suspected false detection data segments. The sum of the probabilities of all suspected false detection segments being identified as physiological states and the number of suspected false detection segments are then calculated, and the self-assessment adjustment coefficient is generated based on both.

[0072] In step S130, if the triggering condition is user feedback triggering, logical consistency and reliability assessment are performed based on the physiological state detection result, the physiological parameters and user feedback information to obtain the feedback confidence level.

[0073] Optionally, the physiological state detection results are objective judgments output by the device-side model (e.g., three snoring sounds detected between 23:00 and 01:00 the previous night, with corresponding probabilities of 0.85, 0.92, and 0.88, respectively). User feedback information is a subjective correction submitted by the user based on their own experience (e.g., information from a bed partner, or verification through audio recordings) (e.g., only one genuine snoring sound was detected during this period, and the other two were false detections of environmental noise). Assuming the user feedback is accurate, the detected physiological state type is adjusted (based on the detection probability, the first and third are adjusted to suspected false snoring sounds). The adjusted physiological state type is then evaluated for logical consistency and reliability to assess whether the feedback information is reasonable and reliable, thus obtaining the feedback confidence level. A higher feedback confidence level indicates stronger reliability of the feedback information and a greater impact of user feedback during subsequent model parameter updates. Conversely, a lower confidence level indicates lower reliability of the feedback information and a possibility of feedback distortion, with a smaller impact of user feedback during subsequent model parameter updates. This confidence level is ultimately correlated with the self-assessment adjustment coefficient and the historical feedback information coefficient.

[0074] The logical consistency and reliability assessment includes: detecting whether the duration of the physiological state meets preset requirements (e.g., whether the snoring time exceeds 12 seconds), and whether the user is in a specific state (e.g., the snoring requires the user to be in bed and asleep).

[0075] Specifically, based on the physiological state detection results, the number of physiological states (such as snoring) detected and their corresponding probabilities for the corresponding time period are obtained. Then, based on user feedback, the number of physiological state detections and their corresponding probabilities are estimated to obtain the actual number of physiological states for the corresponding time period. Segments in user feedback that are inconsistent with the detection results (such as periods where snoring is detected but users report as environmental noise) are marked as suspected physiological state data segments. Then, the logical consistency and reliability of all suspected physiological state data segments can be evaluated to obtain the number of suspected false positive physiological state segments reported by users. Finally, the feedback confidence level is calculated based on the actual number of physiological states and the number of suspected false positive physiological state segments reported by users.

[0076] In step S140, the model parameters of the preset physiological state detection model are updated based on the self-assessment adjustment coefficient, feedback confidence level, and historical feedback information coefficient. The historical feedback information coefficient is determined by the trend and rate of change of the historical feedback detection deviation level.

[0077] Optionally, firstly, compare the difference between the number of times the user reports the actual physiological state and the number of times the model detects it within the target time period, and divide the single feedback detection deviation level according to the piecewise function; then, perform linear fitting on multiple historical feedback detection deviation levels, and use the slope to judge the long-term trend; then, calculate the difference between the previous and current feedback detection deviation levels, and combine it with the feedback interval to obtain the short-term rate of change; finally, fuse the trend and rate according to the preset weights, substitute them into the exponential function mapping, and the historical feedback information coefficient can be obtained.

[0078] Finally, based on the self-evaluation adjustment coefficient, feedback confidence level, and historical feedback information coefficient calculated above, the model parameters of the preset physiological state detection model are updated. Specifically, the parameter update step size can be calculated first. By dynamically adjusting the parameter update step size, the parameter updates are adapted to the needs of the scenario. Then, the updated physiological state feature array and its corresponding labels are obtained. These labels can be obtained based on the triggering conditions (e.g., if triggered by user feedback, the labels are determined based on user correction results; if triggered autonomously by the model, the labels are determined based on valid detection results after self-evaluation). Next, historical physiological state feature data and their corresponding labels, along with the updated physiological state feature array and its corresponding labels, are combined to form a parameter update dataset. Finally, the parameter update dataset is randomly ordered and loaded into the deep model for parameter forward / backward propagation to adjust the model parameters. The learning rate during the backward propagation process is the parameter update step size. The loss function may include the cross-entropy loss function, the optimizer may be stochastic gradient descent (SGD), and the data size of each training batch may be 256 (the specific size can be adjusted according to the server's computing power to ensure a balance between training efficiency and model convergence).

[0079] It should be noted that the calculation process for the model parameter update step size is as follows: First, a pre-set base update step size can be determined. Then, by combining historical feedback-related information coefficients and the adjustment relationships used to distinguish the influence weights of different triggering conditions (such as whether they are triggered by user feedback), these factors are integrated to adjust the base step size, resulting in the final step size actually used in this model parameter update process. The purpose is to ensure that the magnitude of the parameter update is adapted to the needs of the current model optimization scenario, allowing for more reasonable adjustment of model parameters, whether user feedback is involved or the update is based on detection results triggered by the model itself. The specific model parameter update step size can be calculated using the following formula:

[0080] ;

[0081] in, The preset parameter update step size can take the following values: , This is a self-assessment adjustment coefficient. To provide feedback on confidence level, This represents the coefficient for historical feedback information.

[0082] This application embodiment provides a model parameter update method, including: when a preset physiological state detection model parameter update trigger condition is met, acquiring physiological state detection results and physiological parameters for a target object within a target time period, wherein the trigger condition includes user feedback triggering and model self-triggering; evaluating the physiological state detection results based on the physiological parameters to generate a self-evaluation adjustment coefficient; if the trigger condition is user feedback triggering, performing logical consistency and reliability evaluation based on the physiological state detection results, the physiological parameters, and user feedback information; updating the model parameters of the preset physiological state detection model according to the self-evaluation adjustment coefficient, feedback confidence, and historical feedback information coefficient, wherein the historical feedback information coefficient is determined by the historical feedback detection deviation level change trend and change rate; and updating the model parameters of the preset physiological state detection model according to the self-evaluation adjustment coefficient, feedback confidence, and historical feedback information coefficient. In this embodiment, on the one hand, by combining user feedback triggering and model self-triggering for update, it can respond instantly when the user discovers obvious misjudgments, quickly correct sudden deviations, and actively capture scene changes or long-term data patterns. It can complete model iteration without manual intervention, which greatly improves the model's adaptability to complex sleep scenarios and individual user differences, and significantly reduces the probability of false detection and missed detection. On the other hand, by quantifying the fit between physiological parameters and detection results through self-evaluation adjustment coefficients, associating the matching between user feedback and detection results with feedback confidence, and fusing long-term deviation trends and short-term change rates with historical feedback information coefficients, the three factors work together to calculate the model parameter update step size and update value. This ensures that parameter adjustments not only fit the physiological logic of the current scenario but also retain historical effective monitoring experience, avoiding over-updates or under-updates. At the same time, the dynamic update mechanism can also adapt to the long-term fluctuations of user physiological characteristics (such as changes in snoring frequency) and sensor performance (such as sensitivity decay), enhancing model robustness and enabling the device to output stable and accurate detection results in all scenarios and throughout its entire life cycle.

[0083] In one embodiment of this application, after evaluating the physiological state detection results based on the physiological parameters and generating a self-evaluation adjustment coefficient, the method further includes:

[0084] If the triggering condition is model-initiated, the model parameters of the preset physiological state detection model are updated based on the self-evaluation adjustment coefficient and the historical feedback information coefficient.

[0085] Optionally, firstly, compare the difference between the number of times the user confirms the actual physiological state and the number of times the model detects it within the target time period, and divide the single feedback detection deviation level according to the piecewise function; then, perform linear fitting on multiple historical feedback detection deviation levels, and use the slope to judge the long-term trend; then, calculate the difference between the previous and current feedback detection deviation levels, and combine it with the short-term change rate of the feedback interval; finally, fuse the trend and rate according to the preset weights, and substitute them into the exponential function mapping to obtain the historical feedback information coefficient.

[0086] Finally, based on the self-evaluation adjustment coefficient and historical feedback information coefficient calculated above, the model parameters of the preset physiological state detection model are updated. Specifically, the parameter update step size can be calculated first. By dynamically adjusting the parameter update step size, the parameter updates are adapted to the scenario requirements. Then, the updated physiological state feature array and its corresponding labels are obtained. These labels can be determined based on the effective detection results after filtering by a self-evaluation adjustment coefficient (e.g., removing segments labeled 'suspected false detection' and retaining segments determined to be true physiological states). Next, historical physiological state feature data and their corresponding labels, along with the updated physiological state feature array and its corresponding labels, are combined to form a parameter update dataset. Finally, the parameter update dataset is randomly ordered and loaded into a preset physiological state detection model for parameter forward / backward propagation to adjust the model parameters. The learning rate of the backward propagation process is the parameter update step size. The loss function can include the cross-entropy loss function, the optimizer can be stochastic gradient descent (SGD), and the data size of each training batch can be 256 (the specific size can be adjusted according to the server's computing power to ensure a balance between training efficiency and model convergence).

[0087] It should be noted that the calculation process for the model parameter update step size is as follows: First, a pre-set base update step size can be determined. Then, by incorporating relevant information coefficients from historical feedback, the base step size is adjusted to obtain the final step size actually used in this model parameter update process. This model parameter update step size can be calculated using the following formula:

[0088] ;

[0089] in, The preset parameter update step size can take the following values: , This is a self-assessment adjustment coefficient. This represents the coefficient for historical feedback information.

[0090] In one embodiment of this application, the step of performing logical consistency and reliability assessment based on the physiological state detection results, the physiological parameters, and user feedback information includes:

[0091] Based on the physiological state detection results, the number of physiological state events detected and their corresponding probabilities for the target time period are obtained.

[0092] Based on the user feedback information, the number of actual physiological state events occurring within the target time period is estimated, and suspected physiological state data fragments are marked.

[0093] All suspected physiological state data segments are evaluated to obtain the number of suspected false physiological state segments reported by users;

[0094] The confidence level of the feedback is obtained based on the number of false positives for suspected physiological states reported by users and the number of times actual physiological state events occur.

[0095] Optionally, when it is necessary to determine whether the model parameter update is triggered by user feedback, the feedback confidence is obtained by combining user feedback information, the characteristics of physiological state detection in the corresponding time period, and the output state information. First, the number of physiological state events detected and their corresponding probabilities for the target time period can be extracted from the physiological state detection results pre-uploaded by the device. Then, the actual number of occurrences of physiological states is estimated based on user feedback information, and suspected physiological state data segments are marked. Next, logical consistency and reliability assessments are performed on these suspected segments, such as finding all detection results that are judged as the target physiological state (e.g., snoring) and their duration. Data segments with a duration less than a preset threshold TH8 (e.g., 12 seconds) are marked as suspected false detections. Then, the remaining data segments are judged for the target object's state such as being out of bed or in a non-sleep state, thereby finding all suspected false detection data segments as the number of suspected physiological state false detection segments in user feedback. Subsequently, the suspected physiological state segment labels are changed to physiological state, and the false detection segment labels are changed to non-physiological state to form a second physiological state detection result. Finally, the feedback confidence is calculated based on this to measure the credibility of the physiological state detection results after user feedback correction, supporting the subsequent model parameter update process.

[0096] Specifically, user feedback information can include any of the following: severe missed snoring, missed snoring, no missed snoring, severe false snoring, false snoring, or no false snoring. Different user feedback information can correspond to different prediction rules for estimating the frequency of occurrence of the actual physiological state and marking suspected physiological state data segments.

[0097] For example, taking snoring as a physiological state, the prediction rule is explained as follows:

[0098] If the user does not report any missed or false snoring periods, the actual number of snoring sounds during the corresponding period will be estimated. The rules are as follows:

[0099] If the feedback status indicates severe missed snoring, then Number of snoring detections *3, if If it is 0, then The default value is TH12, for example, 100;

[0100] If the feedback status is "snoring missed", then for *1.25, if If it is 0, then The default value is TH13, for example, 50;

[0101] If the feedback status indicates severe false snoring detection, then for ;

[0102] If the feedback status is a false snoring detection, then for ;

[0103] In scenarios without time-based feedback, the method for marking suspected snoring segments is as follows:

[0104] The snoring detection probabilities within the corresponding time period were sorted in descending order, and the top [probabilities] were selected. Snoring data segments are marked as suspected snoring data segments;

[0105] If the user reports a period of missed or false snoring, the estimated number of actual snoring sounds during that period will be calculated. The rules are as follows:

[0106] If the feedback status indicates severe missed snoring, then The result is the number of snoring detected plus the number of snoring detected during the missed detection period multiplied by 3. If the number of snoring detected during the missed detection period is 0, then... The number of snoring sounds detected plus a fixed threshold TH12;

[0107] If the feedback status is "snoring missed", then The result is the number of snoring detected plus the number of snoring detected during the missed detection period multiplied by 1.25. If the number of snoring detected during the missed detection period is 0, then... The number of snoring detections plus a fixed threshold TH13;

[0108] If the feedback status indicates severe false snoring detection, then Number of snoring detections minus number of false positives within a given time period;

[0109] If the feedback status is a false snoring detection, then The formula is: Number of snoring detections - Number of false positives during the specified time period / 1.5;

[0110] If there is feedback information for a time period, the method for marking suspected snoring segments is as follows: sort the snoring detection probability in descending order within the suspected missed or false detection time period, select the snoring data segment with the number of pre-corrected detections, and mark it as a suspected snoring data segment.

[0111] Among them, TH12 and TH13 can be set based on the average number of snoring sounds of similar users in history (or based on the minimum effective detection level of the device).

[0112] In one embodiment of this application, updating the model parameters of the preset physiological state detection model based on the self-assessment adjustment coefficient, feedback confidence level, and historical feedback information coefficient includes:

[0113] The labels of suspected physiological state data segments are changed to physiological state labels, and the labels of suspected false physiological state data segments are changed to non-physiological state labels to form the first physiological state detection result.

[0114] Based on the first physiological state detection result, obtain the updated physiological state feature array and corresponding labels;

[0115] Obtain the historical physiological state feature array and corresponding labels, and combine it with the updated physiological state feature array and corresponding labels to update the dataset as parameters;

[0116] The dataset is updated based on the parameters, and the preset physiological state detection model is iteratively trained to obtain the updated preset physiological state detection model.

[0117] Optionally, based on the suspected physiological state data fragments corrected by user feedback (i.e., fragments identified as missed detections in user feedback) and suspected physiological state false detection fragments (i.e., fragments identified as false detections in user feedback), the former's label is changed to a physiological state label, and the latter's label is changed to a non-physiological state label, forming a first physiological state detection result. Then, based on this first detection result, an updated physiological state feature array containing the latest annotation information and its corresponding labels are extracted and obtained. Subsequently, previously stored historical physiological state feature arrays and corresponding labels are retrieved and integrated with the updated feature array and labels to construct a complete parameter update dataset for model parameter updates. Finally, the parameter update dataset is randomly sorted (to avoid interference from data temporal sequence on model training and ensure uniform distribution of training samples) and loaded into a preset physiological state detection model for parameter forward / backward propagation to adjust model parameters. The learning rate during backward propagation is the parameter update step size. The loss function can include the cross-entropy loss function, the optimizer can be stochastic gradient descent (SGD), and the amount of data in each training batch can be 256. By continuously optimizing the model parameters to adapt to the combined patterns of new and historical data, the pre-set physiological state detection model with updated parameters is finally obtained, achieving iterative improvement in the model's detection accuracy.

[0118] In one embodiment of this application, before updating the model parameters of the preset physiological state detection model based on the self-assessment adjustment coefficient, feedback confidence level, and historical feedback information coefficient, the following steps are included:

[0119] Based on the difference between the number of times the actual physiological state occurs and the number of physiological state events detected within the target time period, the feedback detection bias level is determined by a piecewise function.

[0120] The historical feedback detection deviation level is linearly fitted, and the slope of the fitted line is used as the trend of the historical feedback detection deviation level.

[0121] The rate of change is obtained based on the deviation level of the previous historical feedback detection and the deviation level of the current feedback detection;

[0122] The historical feedback detection deviation level change trend and change rate are weighted and fused, and then mapped through an exponential function to obtain the historical feedback information coefficient.

[0123] Optionally, first compare the difference between the number of times the user confirms the actual physiological state and the number of times the model detects it within the target time period, and divide the single feedback detection deviation level according to the piecewise function; then perform linear fitting on multiple historical feedback detection deviation levels, and use the slope to judge the long-term trend; then calculate the difference between the previous and current feedback detection deviation levels, and combine it with the feedback interval to obtain the short-term rate of change; finally, fuse the trend and rate according to the preset weights, and substitute them into the exponential function mapping to obtain the historical feedback information coefficient.

[0124] The classification of single-feedback detection deviation levels according to piecewise functions can be expressed by the following formula:

[0125] ;

[0126] Wherein, the horizontal axis of the fitted data represents the time difference between the historical feedback information and the first feedback information. This is the weighting coefficient, which can be 0.5. The change trend of the deviation level of the previous historical feedback detection and the deviation level of the current feedback detection is calculated by dividing the difference between the current deviation level and the previous feedback deviation level by the time interval between the two feedbacks, where the time interval is in days.

[0127] Taking snoring detection as an example, the deviation levels of snoring detection obtained from the above calculations are analyzed. After fitting, the slope of the fitted line is... As a trend of historical feedback detection deviation level changes, the historical feedback information coefficient It can be expressed by the following formula:

[0128] ;

[0129] Where e is the natural constant, The trend of historical feedback snoring deviation levels is obtained by linearly fitting the deviation of the number of historical feedback snoring detections and calculating its slope. The feedback snoring detection deviation level for The difference between the number of snoring detected and the corresponding time period Corresponding value.

[0130] In one embodiment of this application, the step of evaluating the physiological state detection results based on the physiological parameters and generating a self-evaluation adjustment coefficient includes:

[0131] In the physiological state detection results, find all detection results that are determined to be physiological states and the duration of the physiological states;

[0132] Physiological state data segments whose duration is less than a preset time threshold are marked as suspected false physiological state segments.

[0133] Based on the physiological parameters, the remaining detection results that are determined to be physiological states are verified, and physiological state data segments that meet the preset false detection conditions are marked as suspected false detection physiological state segments.

[0134] Determine the number of all false positives labeled as suspected physiological states and the sum of the probabilities of all false positives labeled as suspected physiological states being identified as physiological states, and calculate the self-evaluation adjustment coefficient.

[0135] Optionally, in the physiological state detection results, all detection results judged as the target physiological state (such as snoring) and their duration are found. Data segments with a duration less than a preset threshold TH8 (e.g., 12 seconds) are marked as suspected false detections. Then, the remaining data segments are judged for the target object's state, such as being out of bed or not asleep, to find all suspected false detection data segments, which are taken as the number of suspected physiological state false detection segments reported by the user. Then, the sum of the probabilities of all suspected physiological state false detection segments being judged as physiological states by the model is calculated. and the number of suspected false positives for physiological states Each of these two values ​​is compared with its corresponding threshold (e.g., the threshold coefficient for the quantity is set to 750, and the threshold coefficient for the probability sum is set to 600), resulting in two ratios. The larger of the two ratios is then taken. Next, this larger value is compared with the self-assessment adjustment coefficient threshold (e.g., set to 0.2), and the smaller of the two values ​​is taken. This smaller value is the self-assessment adjustment coefficient. It is used to measure the self-evaluation of the model's detection results and to assist in subsequent processes such as updating model parameters.

[0136] Self-assessment adjustment coefficient Specifically, it is calculated using the following formula:

[0137] ;

[0138] Among them, TH9 is The corresponding threshold coefficients are 750 for TH9 and TH10. The corresponding threshold coefficients are TH10 (which can be 600) and TH11 (which is the self-evaluation adjustment coefficient threshold, which can be 0.2).

[0139] It should be noted that the labels of suspected false physiological state segments can be changed to non-physiological states to form a second physiological state detection result, which can be used for subsequent model parameter update operations in autonomously triggered states.

[0140] In one embodiment of this application, updating the model parameters of the preset physiological state detection model based on the self-assessment adjustment coefficient, feedback confidence level, and historical feedback information coefficient includes:

[0141] If the triggering condition is triggered autonomously by the model, then all labels marked as suspected false physiological states will be modified to non-physiological states, forming a second physiological state detection result;

[0142] Based on the second physiological state detection result, obtain the updated physiological state feature array and corresponding labels;

[0143] Obtain the historical physiological state feature array and corresponding labels, and combine it with the updated physiological state feature array and corresponding labels to update the dataset as parameters;

[0144] The preset physiological state detection model is iteratively trained based on the updated dataset of the parameters to obtain the updated preset physiological state detection model.

[0145] Optionally, if the triggering condition is triggered autonomously by the model, based on the suspected false physiological state segments marked during the generation of the self-evaluation adjustment coefficient (such as segments with a duration less than the threshold or that contradict the bed-off state), their labels are modified to non-physiological states, forming a second physiological state detection result. Then, based on this second physiological state detection result, an updated physiological state feature array containing the latest annotation information and its corresponding labels are extracted and obtained. Subsequently, the previously stored historical physiological state feature arrays and corresponding labels are retrieved and integrated with the updated feature arrays and labels to construct a complete parameter update dataset for model parameter updates. Finally, the parameter update dataset is randomly arranged (to avoid the data temporal sequence causing the model to overfit the features of a certain time period and to ensure that the training samples are evenly distributed) and loaded into the preset physiological state detection model for parameter forward / backward propagation to adjust the model parameters. The learning rate of the backward propagation process is the parameter update step size. The loss function can include the cross-entropy loss function, the optimizer can be stochastic gradient descent (SGD), and the amount of data in each training batch can be 256. By continuously optimizing the model parameters to adapt to the combined patterns of new and historical data, the pre-set physiological state detection model with updated parameters is finally obtained, achieving iterative improvement in the model's detection accuracy.

[0146] like Figure 3 As shown in one embodiment of this application, after updating the model parameters of the preset physiological state detection model based on the self-assessment adjustment coefficient, feedback confidence level, and historical feedback information coefficient, the method further includes:

[0147] The dataset is updated based on the parameters, and the updated preset physiological state detection model is evaluated to obtain the number of suspected false physiological state segments.

[0148] If the number of suspected false physiological state segments is less than a preset threshold, the historical physiological state feature dataset is updated.

[0149] If the number of suspected false physiological state segments is not less than the preset threshold, the model parameter update step size is adjusted.

[0150] If the number of times the model parameter update step size is adjusted is less than a preset threshold, the model is retrained based on the updated parameter dataset, and the number of suspected physiological state false detection segments in the updated model is evaluated.

[0151] If the number of times the model parameter update step size is adjusted is not less than the preset threshold, the update fails and the original model parameters are retained.

[0152] Optionally, after completing the construction of the parameter update dataset and iteratively training the model, self-evaluation is performed on the updated physiological state feature array and historical feature data respectively. The specific process is as follows: For the updated feature array, a new detection result needs to be generated: First, for the updated physiological state feature array, a self-evaluation is performed according to a specific process. That is, for the updated physiological state feature array, a logical consistency and reliability self-evaluation can be performed (find all detection results that are judged as the target physiological state (such as snoring) and their duration. If the duration is less than the preset threshold TH8 (such as 12 seconds), the data segment is marked as a suspected false detection. Then, the remaining data segments are judged for the target object's state such as getting out of bed or in a non-sleep state, thereby finding all suspected false detection data segments. The label of the suspected physiological state false detection segment is changed to a non-physiological state to form a new physiological state detection result).

[0153] For historical feature data, only suspected false positive segments are marked without modifying the labels. This is done by finding all detection results identified as the target physiological state (e.g., snoring) and their duration. Data segments with a duration less than a preset threshold TH8 (e.g., 12 seconds) are marked as suspected false positives. Then, the remaining data segments are analyzed to determine the target object's state (e.g., whether it's out of bed or asleep), thus identifying all suspected false positive data segments.

[0154] Based on the above self-evaluation results, obtain the proportion of suspected physiological state false detection segments in the parameter update dataset to the total number of segments that the model classifies as physiological states. ,like If the value is less than a preset threshold TH14 (TH14 being, for example, 0.05), indicating that the self-assessment result meets the preset assessment requirements, then the updated model parameters are saved, and the historical physiological state feature dataset is updated. The update method can be to save snoring feature data and physiological features with a probability of physiological state higher than the preset threshold TH6 or lower than the preset threshold TH7 to the historical physiological state feature dataset, and delete physiological state feature data with a probability of physiological state higher than the preset threshold TH6 or lower than the preset threshold TH7 that do not meet the updated model prediction results from the historical physiological state feature dataset. Here, TH6 is the high confidence threshold for determining a physiological state (e.g., 0.8), and TH7 is the high confidence threshold for determining a non-physiological state (e.g., 0.2).

[0155] Otherwise, adjust the model parameter update step size, incrementing the number of adjustments. The method for adjusting the parameter update step size can refer to the learning rate update optimization process of adaptive learning rate optimizers (such as Adam and RMSProp). Then, determine if the number of parameter update step size adjustments is less than a preset threshold TH15 (TH15 is, for example, 5). If so, retrain the model based on the adjusted parameter update step size, and then perform a self-evaluation process on the newly trained model. Otherwise, the parameter update fails, and the original model parameters are saved.

[0156] Finally, the latest parameters of the preset physiological state detection model can be synchronized and updated to the device. If the model parameters are updated successfully, the updated model parameters from the server will be synchronized to the device; if the parameter update fails, no operation will be performed.

[0157] In this embodiment, on the one hand, by combining user feedback triggering and model self-triggering for update, it can respond instantly when the user discovers obvious misjudgments, quickly correct sudden deviations, and actively capture scene changes or long-term data patterns. It can complete model iteration without manual intervention, which greatly improves the model's adaptability to complex sleep scenarios and individual user differences, and significantly reduces the probability of false detection and missed detection. On the other hand, by quantifying the fit between physiological parameters and detection results through self-evaluation adjustment coefficients, associating the matching between user feedback and detection results with feedback confidence, and fusing long-term deviation trends and short-term change rates with historical feedback information coefficients, the three factors work together to calculate the model parameter update step size and update value. This ensures that parameter adjustments not only fit the physiological logic of the current scenario but also retain historical effective monitoring experience, avoiding over-updates or under-updates. At the same time, the dynamic update mechanism can also adapt to the long-term fluctuations of user physiological characteristics (such as changes in snoring frequency) and sensor performance (such as sensitivity decay), enhancing model robustness and enabling the device to output stable and accurate detection results in all scenarios and throughout its entire life cycle.

[0158] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0159] Example 2

[0160] In one embodiment, a model parameter updating device is provided, which corresponds one-to-one with the model parameter updating method in the above embodiments. For example... Figure 4 As shown, the model parameter update includes a data acquisition unit 10, an evaluation unit 20, a feedback confidence acquisition unit 30, and a model parameter update unit 40. Detailed descriptions of each functional module are as follows:

[0161] The data acquisition unit 10 is used to acquire the physiological state detection results and physiological parameters of the target object within a target time period when the preset triggering conditions for updating the physiological state detection model parameters are met. The triggering conditions include user feedback triggering and model autonomous triggering.

[0162] Evaluation unit 20 is used to evaluate the physiological state detection results based on the physiological parameters and generate a self-evaluation adjustment coefficient;

[0163] The feedback confidence acquisition unit 30 is used to perform logical consistency and reliability assessment based on the physiological state detection result, the physiological parameters and the user feedback information if the triggering condition is user feedback triggering.

[0164] The model parameter update unit 40 is used to update the model parameters of the preset physiological state detection model based on the self-evaluation adjustment coefficient, feedback confidence level and historical feedback information coefficient, wherein the historical feedback information coefficient is determined by the trend and rate of change of historical feedback detection deviation level.

[0165] In one embodiment of this application, the model parameter update unit 40 is further configured to:

[0166] If the triggering condition is model-initiated, the model parameters of the preset physiological state detection model are updated based on the self-evaluation adjustment coefficient and the historical feedback information coefficient.

[0167] In one embodiment of this application, the feedback confidence acquisition unit 30 is further configured to:

[0168] Based on the physiological state detection results, the number of physiological state events detected and their corresponding probabilities for the target time period are obtained.

[0169] Based on the user feedback information, the number of actual physiological state events occurring within the target time period is estimated, and suspected physiological state data fragments are marked.

[0170] All suspected physiological state data segments are evaluated to obtain the number of suspected false physiological state segments reported by users;

[0171] The confidence level of the feedback is obtained based on the number of false positives for suspected physiological states reported by users and the number of times actual physiological state events occur.

[0172] In one embodiment of this application, the model parameter update unit 40 is further configured to:

[0173] The labels of suspected physiological state data segments are changed to physiological state labels, and the labels of suspected false physiological state data segments are changed to non-physiological state labels to form the first physiological state detection result.

[0174] Based on the first physiological state detection result, obtain the updated physiological state feature array and corresponding labels;

[0175] Obtain the historical physiological state feature array and corresponding labels, and combine it with the updated physiological state feature array and corresponding labels to update the dataset as parameters;

[0176] The dataset is updated based on the parameters, and the preset physiological state detection model is iteratively trained to obtain the updated preset physiological state detection model.

[0177] In one embodiment of this application, the device further includes: a historical feedback information coefficient acquisition unit, used for:

[0178] Based on the difference between the number of times the actual physiological state occurred and the number of physiological state events detected within the target time period, the historical feedback detection bias level is determined by a piecewise function.

[0179] The historical feedback detection deviation level is linearly fitted, and the slope of the fitted line is used as the trend of the historical feedback detection deviation level.

[0180] The rate of change is obtained based on the difference between the previous historical feedback detection deviation level and the feedback detection deviation level;

[0181] The historical feedback detection deviation level change trend and change rate are weighted and fused, and then mapped through an exponential function to obtain the historical feedback information coefficient.

[0182] In one embodiment of this application, the evaluation unit 20 is further configured to:

[0183] In the physiological state detection results, find all detection results that are determined to be physiological states and the duration of the physiological states;

[0184] Physiological state data segments whose duration is less than a preset time threshold are marked as suspected false physiological state segments.

[0185] Based on the physiological parameters, the remaining detection results that are determined to be physiological states are verified, and physiological state data segments that meet the preset false detection conditions are marked as suspected false detection physiological state segments.

[0186] Determine the number of all false positives labeled as suspected physiological states and the sum of the probabilities of all false positives labeled as suspected physiological states being identified as physiological states, and calculate the self-evaluation adjustment coefficient.

[0187] In one embodiment of this application, the model parameter update unit 40 is further configured to:

[0188] If the triggering condition is triggered autonomously by the model, then all labels marked as suspected false physiological states will be modified to non-physiological states, forming a second physiological state detection result;

[0189] Based on the second physiological state detection result, obtain the updated physiological state feature array and corresponding labels;

[0190] Obtain the historical physiological state feature array and corresponding labels, and combine it with the updated physiological state feature array and corresponding labels to update the dataset as parameters;

[0191] The preset physiological state detection model is iteratively trained based on the updated dataset of the parameters to obtain the updated preset physiological state detection model.

[0192] In one embodiment of this application, the apparatus further includes: a model evaluation unit, which is further configured to:

[0193] The dataset is updated based on the parameters, and the updated preset physiological state detection model is evaluated to obtain the number of suspected false physiological state segments.

[0194] If the number of suspected false physiological state segments is less than a preset threshold, the historical physiological state feature dataset is updated.

[0195] If the number of suspected false physiological state segments is not less than the preset threshold, the model parameter update step size is adjusted.

[0196] If the number of times the model parameter update step size is adjusted is less than a preset threshold, the model is retrained based on the updated parameter dataset, and the number of suspected physiological state false detection segments in the updated model is evaluated.

[0197] If the number of times the model parameter update step size is adjusted is not less than the preset threshold, the update fails and the original model parameters are retained.

[0198] In this embodiment, on the one hand, by combining user feedback triggering and model self-triggering for update, it can respond instantly when the user discovers obvious misjudgments, quickly correct sudden deviations, and actively capture scene changes or long-term data patterns. It can complete model iteration without manual intervention, which greatly improves the model's adaptability to complex sleep scenarios and individual user differences, and significantly reduces the probability of false detection and missed detection. On the other hand, by quantifying the fit between physiological parameters and detection results through self-evaluation adjustment coefficients, associating the matching between user feedback and detection results with feedback confidence, and fusing long-term deviation trends and short-term change rates with historical feedback information coefficients, the three factors work together to calculate the model parameter update step size and update value. This ensures that parameter adjustments not only fit the physiological logic of the current scenario but also retain historical effective monitoring experience, avoiding over-updates or under-updates. At the same time, the dynamic update mechanism can also adapt to the long-term fluctuations of user physiological characteristics (such as changes in snoring frequency) and sensor performance (such as sensitivity decay), enhancing model robustness and enabling the device to output stable and accurate detection results in all scenarios and throughout its entire life cycle.

[0199] Specific limitations regarding the model parameter update device can be found in the limitations of the model parameter update method described above, and will not be repeated here. Each module in the aforementioned model parameter update device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0200] Example 3

[0201] In one embodiment, a model parameter update system is provided, which includes the model parameter update device described in the above embodiment. For example... Figure 1 As shown, the model parameter update system includes device S1, client S2, and server S3. Detailed descriptions of each functional module are as follows:

[0202] The device S1 is used to extract features from the acquired physiological data of the target object and input the extracted features into a preset physiological state detection model for prediction to obtain the physiological state detection result.

[0203] Client S2 is used to display the physiological state detection results to the user. If the user has feedback on the physiological state detection results, the user can input feedback information.

[0204] Server S3 includes the model parameter update device described in the second embodiment above, used to acquire physiological state detection results and physiological parameters of a target object within a target time period when the triggering conditions for updating preset physiological state detection model parameters are met, wherein the triggering conditions include user feedback triggering and model self-triggering; evaluate the physiological state detection results based on the physiological parameters to generate a self-evaluation adjustment coefficient; if the triggering condition is user feedback triggering, perform logical consistency and reliability evaluation based on the physiological state detection results, the physiological parameters, and user feedback information; update the model parameters of the preset physiological state detection model according to the self-evaluation adjustment coefficient, feedback confidence level, and historical feedback information coefficient, wherein the historical feedback information coefficient is determined by the trend and rate of change of historical feedback detection deviation level.

[0205] Optionally, the client S2 includes, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. Specifically, it may include a physiological state data display module and a physiological state detection feedback module. The physiological state data display module refers to an interface within the client showing the user's physiological state statistics from the previous night. Taking snoring as an example, it could be a bar chart showing the number of snoring sounds at equal time intervals during sleep, with each time interval potentially being 10 minutes. The snoring detection feedback module involves setting up a shortcut interface within the physiological state data display interface. Triggering this shortcut leads to the physiological state detection feedback interface, where users can provide feedback on two aspects: missed physiological state detections and false positives. Feedback information may include any one of the following: severe missed detections, missed detections, no missed detections, severe false positives, false positives, and no false positives.

[0206] Specifically, taking snoring as an example, the main manifestation of severe missed detection is that the number of snoring detected is less than the number of snoring sounds perceived by the user * 0.3. The main manifestation of missed snoring is that the number of snoring detected is less than the number of snoring sounds perceived by the user * 0.8 but greater than the number of snoring sounds perceived by the user * 0.3. The main manifestation of no missed detection is that the number of snoring detected is not less than the number of snoring sounds perceived by the user * 0.8. The main manifestation of severe false snoring is that the algorithm detects snoring during a period when the user does not snore, and the corresponding period lasts for more than 1 hour. The main manifestation of false snoring is that the algorithm detects snoring during a period when the user does not snore, and the corresponding period lasts for more than 30 minutes but less than 1 hour. The main manifestation of no false snoring is that the algorithm detects snoring during a period when the user does not snore, and the corresponding period lasts for less than 30 minutes. Other feedback information includes: missed detection time segments and false detection time segments. The snoring detection feedback module also includes an introduction to the judgment methods for missed and false snoring, as well as the factors affecting the accuracy of snoring detection.

[0207] Furthermore, the S3 server can be implemented using a standalone server or a server cluster consisting of multiple servers. It may include a physiological state data storage module and a preset physiological state detection model parameter update module. The physiological state data storage module stores historical physiological state feature data, physiological state detection results, physiological parameter data, and historical physiological state detection feedback data. Historical physiological state feature data refers to physiological state feature data that meets preset requirements (including physiological state feature data whose probability of being judged as a physiological state by the current physiological state detection model is higher than a preset threshold TH6 or lower than a preset threshold TH7, where TH6 can be 0.8 and TH7 can be 0.2; after each model update, historical feature data that does not meet the above requirements will be removed. To reduce the computational load and storage space caused by historical data updates, after the amount of physiological state feature data reaches its upper limit, earlier physiological state feature data in the time dimension will be removed). The physiological state detection result is the result of the current physiological state detection model's processing and judgment of the historical physiological state feature data, including… The model includes whether a state is physiological and its corresponding probability. The physiological parameters are physiological parameter feature data corresponding to the time period of the physiological state feature data. These physiological parameter feature data include the out-of-bed state and the sleep state. The physiological parameters are obtained through airbag pressure values ​​or matrix pressure sensor signals. The historical physiological state detection feedback data includes user feedback status and differences in snoring detection results within the feedback time period before and after the model update. The differences in physiological state detection results include the number of changes in detection results before and after the model update, the number of states judged as physiological from non-physiological states, the number of states judged as non-physiological from physiological states, and the ratio of the change in the number of detected physiological states to the number of detected physiological states before the model update. The physiological state detection model parameter update module includes model update and model self-evaluation. The model update includes feedback update and autonomous update, where feedback update is triggered by the user through the snoring detection feedback module, and autonomous update is triggered autonomously by the model under specific conditions.

[0208] In one embodiment of this application, the device includes:

[0209] The signal acquisition module is used to collect physiological data of the target object;

[0210] The signal quality judgment module is used to filter the collected physiological data to obtain target physiological data that meets the quality requirements.

[0211] The feature extraction module is used to extract features from the target physiological data to obtain physiological features;

[0212] The physiological state detection module is pre-configured with the preset physiological state detection model, which is used to predict the physiological characteristics and obtain the physiological state detection results.

[0213] Specifically, the main carrier of the device S1 can be a smart bed frame or smart mattress, etc., and it mainly consists of four modules: a signal acquisition module, a feature extraction module, a signal quality judgment module, and a physiological state detection module. The signal acquisition module is mainly used to collect sensor data through sensors. This sensor data includes physiological data such as sound data and respiratory data. Sound data acquisition methods include, but are not limited to, acquisition using a miniature microphone chip, such as the SPU0410LR5H-QB. Respiratory data acquisition methods include, but are not limited to, acquisition using an airbag pressure sensor or a matrix pressure sensor. The signal quality judgment module is mainly used for signal filtering, eliminating unstable states and signal segments with low signal-to-noise ratios. Unstable states include those within the time period affected by the action and sound signals in an unstable state. The signal sources for action detection include, but are not limited to, airbag pressure signals or matrix pressure sensor signals. The feature extraction module is used to extract features from the physiological data collected by the signal acquisition module after filtering by the signal quality judgment module. The physiological state detection module is used to detect physiological states, such as snoring, sleep apnea events, and sleep states (such as deep sleep, light sleep, etc.). It can use a deep learning model to predict the features extracted by the feature extraction module and output the physiological state detection results. The physiological state detection results can include whether a physiological state has occurred and the probability of the physiological state occurring.

[0214] It should be noted that the signal quality judgment module also includes a signal stability judgment submodule, which is used to determine whether the acquired physiological parameters are in a stable state. Taking the sound signal as an example, the sound signal is in an unstable state when the ratio of the maximum value to the average value of the sound energy in the specified frequency band within the window time TH1 exceeds the preset threshold TH2. For example, TH1 is 12 seconds and TH2 is 20 seconds. Specifically, to determine whether a sound signal is in an unstable state, a window time (e.g., 12 seconds) and a threshold (e.g., 20) must first be set. This is determined by calculating the ratio of the maximum to the average sound energy in a specified frequency band within the window. First, the original sound signal f(t) undergoes first-order differential processing to obtain the processed sound signal f1(t). Then, the differentially processed sound signal f1(t) is bandpass filtered from 80Hz to 1000Hz. Next, the sum of the absolute values ​​of the filtered signal f2(t) from one second prior to the current moment is calculated as the signal energy e(t) at that moment. Finally, the ratio of the maximum to the average energy in the signal frequency band within the window time TH1 is calculated. If this ratio exceeds the threshold TH2, the sound signal is in an unstable state; otherwise, it is in a stable state. The signal-to-noise ratio (SNR) is the ratio snr of the sum of the absolute values ​​of the filtered signal f2(t) to the sum of the absolute values ​​of the original sound signal f(t) within the window time. When this ratio snr is lower than the preset threshold TH3 (e.g., 1), the SNR is considered low.

[0215] In addition, the features extracted by the feature extraction module include, but are not limited to, Mel-frequency cepstral coefficients, Mel-frequency cepstral coefficient differences, nonlinear frequency band energy, high-frequency energy proportion, and the phase difference between the sound signal envelope and the breathing signal. The Mel-frequency cepstral coefficients can be obtained by first pre-emphasizing the original sound signal, as shown in the following formula:

[0216]

[0217] Where t represents the sampling point at time t. High-frequency attenuation compensation coefficient is used to boost high-frequency components and compensate for signal attenuation (e.g., );

[0218] The pre-emphasized signal is framed to obtain a short time window from 40ms before time t to time t itself. For short-time windows After applying a Hamming window, f3(t) is obtained; a Fourier transform is performed on f3(t) to calculate the amplitude spectrum. Calculate the power spectrum corresponding to f3(t). Where N is the number of Fourier transform points; then, the coefficients of the Mel filter bank are obtained. Calculate the energy of each filter bank and take the logarithm. M represents the number of Mel filter banks; feature compression and decorrelation operations are performed to obtain the Mel cepstral coefficients (MFCC coefficients), which can be calculated using the following formula:

[0219] ;

[0220] in, L To retain the number of coefficients, its value is, for example, 12;

[0221] By shifting the sliding window forward by 20ms at time t and repeating the above steps, the Mel-Cepstral Coefficients at different times can be obtained.

[0222] The method for obtaining the Mel-frequency cepstral coefficient difference is as follows:

[0223] ;

[0224] Where K is the difference order;

[0225] The nonlinear frequency band energy acquisition step is similar to that of Mel-frequency cepstral coefficients, including: obtaining a short time window from 40ms before time t to the current time t by performing frame segmentation processing on the pre-emphasized signal. For short-time windows After applying a Hamming window, f3(t) is obtained; a Fourier transform is performed on f3(t) to calculate the amplitude spectrum. Calculate the power spectrum corresponding to f3(t). , where N is the number of Fourier transform points.

[0226] Next, the Bark-scale filter bank coefficients can be obtained. Calculate the energy of each sub-band. , where m is the m-th nonlinear sub-band; the sliding window is shifted 20ms after time t, and the above steps (Mel cepstral coefficient acquisition process) are repeated to obtain the energy characteristics of the nonlinear band at different times, and the number of nonlinear band features is 24.

[0227] The method for calculating the proportion of high-frequency energy is as follows:

[0228] ;

[0229] It should be noted that in the phase difference between the sound signal envelope and the respiratory signal, the sound signal energy envelope is calculated by taking the root mean square value of the original signal f(t) over a preset window time length TH4 to obtain signal f4(t). f4(t) is then passed through a bandpass filter and an all-pass filter bank to obtain the sound signal envelope f5(t). The passband of the aforementioned filter bank is [0.1Hz, 4Hz]. The peak point of f5(t) is then located. The respiratory signal f6(t) is acquired using an airbag pressure sensor or a matrix pressure sensor. f6(t) is then passed through the same bandpass filter and all-pass filter bank to obtain f7(t). The peak point of f7(t) is then located, and the time interval between the peak points of f5(t) and the nearest peak point of f7(t) is calculated. This time interval is then divided by the current average respiratory time interval multiplied by 2. The phase difference between the current sound signal envelope and the respiratory signal is obtained. Before the peak point of f7(t) is found and the time interval between the peak points of f7(t) and f5(t) is less than the current average respiratory interval, the phase difference between the sound signal envelope and the respiratory signal remains unchanged. If it exceeds the current average respiratory interval, the phase difference between the sound signal envelope and the respiratory signal is set to 2π / 2. .

[0230] Furthermore, the physiological state detection module has a pre-set physiological state detection model, which can be a deep learning model. The model input can be an input feature array (9081) formed by splicing together the Mel-frequency cepstral coefficient array (200*12), the Mel-frequency cepstral coefficient difference array (200*12), the nonlinear frequency band energy array (200*24), the high-frequency energy ratio (200*1), and the phase difference between the sound signal envelope and the respiratory signal (1) of all signal segments within the TH5 window time. The model output is the physiological state detection result. Taking snoring as an example, the physiological state detection result includes whether it is snoring and its corresponding probability. The model structure used can be as follows: Figure 5As shown. Input layer 1 corresponds to the Mel-frequency cepstral coefficient array, input layer 2 corresponds to the Mel-frequency cepstral coefficient difference array, input layer 3 corresponds to the nonlinear frequency band energy array, input layer 4 corresponds to the high-frequency energy proportion, and input layer 5 corresponds to the phase difference between the sound signal envelope and the breathing signal; the convolutional layer specifically includes three steps: Conv2d (convolution processing), BN (batch normalization processing), and ReLU (activation processing). K3 represents a convolution kernel size of 3*3. This represents a 3x1 kernel size, s2 a stride of 2, s1 a stride of 1, p1 a padding maxpooling, c1 a single kernel, c16 a 16-kernel configuration, and 200x12x16 representing the number and format of the layer's output parameters. The 1x1 convolutional layer is a 1x1 convolutional layer used to adjust the input size and shape and add non-linearity (including ReLU operations). Fully connected layer 1 is a fully connected layer with ReLU activation, and fully connected layer 2 is the output layer with sigmoid activation. The probability of the input being snoring is calculated; if it exceeds a preset threshold, such as 0.5, the input is considered snoring.

[0231] In one embodiment, a computer device is provided, which may be a terminal device, and its internal structure diagram may be as follows: Figure 6 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computational and control capabilities. The memory includes a readable storage medium storing computer-readable instructions. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer-readable instructions implement a model parameter update method. The readable storage medium provided in this embodiment includes both non-volatile and volatile readable storage media.

[0232] In this application embodiment, a computer device is provided, including a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor. When the processor executes the computer-readable instructions, it implements the steps of the model parameter update method described above.

[0233] In one embodiment of the application, a readable storage medium is provided, which stores computer-readable instructions. When the computer-readable instructions are executed by a processor, they implement the steps of the model parameter update method described above.

[0234] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing related hardware with computer-readable instructions. These computer-readable instructions can be stored in a non-volatile readable storage medium or a volatile readable storage medium. When executed, these computer-readable instructions can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0235] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0236] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for updating model parameters, characterized in that, The method includes: When the preset physiological state detection model parameter update trigger condition is met, the physiological state detection results and physiological parameters of the target object within the target time period are obtained. The trigger condition includes user feedback trigger and model autonomous trigger. The physiological state detection results include whether snoring is detected. The physiological parameters include the original data or basic data of the physiological function state or physiological activity characteristics of the target object in bed state and sleep state. The physiological state detection results are evaluated based on the physiological parameters to generate a self-evaluation adjustment coefficient; If the triggering condition is user feedback trigger, a logical consistency and reliability assessment is performed based on the physiological state detection results, the physiological parameters, and the user feedback information. The model parameters of the preset physiological state detection model are updated based on the self-assessment adjustment coefficient, feedback confidence, and historical feedback information coefficient. The historical feedback information coefficient is determined by the trend and rate of change of historical feedback detection deviation levels. The self-assessment adjustment coefficient is determined based on the sum of probabilities that all suspected physiological state false detection segments are identified as the target physiological state and the number of suspected physiological state false detection segments. The suspected physiological state false detection segments are determined by the duration of the target physiological state, combined with the bed-leaning signal from the pressure sensor and the activity status of the motion sensor. The feedback confidence is determined based on the number of suspected physiological state false detection segments reported by the user and the number of occurrences of actual physiological state events. The number of suspected physiological state false detection segments reported by the user is jointly determined by the physiological state detection results and user feedback information. The target physiological state refers to snoring.

2. The model parameter update method as described in claim 1, characterized in that, After evaluating the physiological state detection results based on the physiological parameters and generating a self-evaluation adjustment coefficient, the method further includes: If the triggering condition is model-initiated, the model parameters of the preset physiological state detection model are updated based on the self-evaluation adjustment coefficient and the historical feedback information coefficient.

3. The model parameter update method as described in claim 1, characterized in that, The logical consistency and reliability assessment based on the physiological state detection results, the physiological parameters, and user feedback information includes: Based on the physiological state detection results, the number of physiological state events detected and their corresponding probabilities for the target time period are obtained. Based on the user feedback information, the number of actual physiological state events occurring within the target time period is estimated, and suspected physiological state data fragments are marked. All suspected physiological state data segments are evaluated to obtain the number of suspected false physiological state segments reported by users; The confidence level of the feedback is obtained based on the number of false positives for suspected physiological states reported by users and the number of times actual physiological state events occur.

4. The model parameter update method as described in claim 3, characterized in that, The step of updating the model parameters of the preset physiological state detection model based on the self-assessment adjustment coefficient, feedback confidence level, and historical feedback information coefficient includes: The labels of suspected physiological state data segments are changed to physiological state labels, and the labels of suspected false physiological state data segments are changed to non-physiological state labels to form the first physiological state detection result. Based on the first physiological state detection result, obtain the updated physiological state feature array and corresponding labels; Obtain the historical physiological state feature array and corresponding labels, and combine it with the updated physiological state feature array and corresponding labels to update the dataset as parameters; The dataset is updated based on the parameters, and the preset physiological state detection model is iteratively trained to obtain the updated preset physiological state detection model.

5. The model parameter update method according to any one of claims 1-4, characterized in that, Before updating the model parameters of the preset physiological state detection model based on the self-assessment adjustment coefficient, feedback confidence level, and historical feedback information coefficient, the following steps are included: Based on the difference between the number of times the actual physiological state occurred and the number of physiological state events detected within the target time period, the historical feedback detection bias level is determined by a piecewise function. The historical feedback detection deviation level is linearly fitted, and the slope of the fitted line is used as the trend of the historical feedback detection deviation level. The rate of change is obtained based on the difference between the previous historical feedback detection deviation level and the feedback detection deviation level; The historical feedback detection deviation level change trend and change rate are weighted and fused, and then mapped through an exponential function to obtain the historical feedback information coefficient.

6. The model parameter update method according to any one of claims 1-4, characterized in that, The step of evaluating the physiological state detection results based on the physiological parameters and generating a self-evaluation adjustment coefficient includes: In the physiological state detection results, find all detection results that are determined to be physiological states and the duration of the physiological states; Physiological state data segments whose duration is less than a preset time threshold are marked as suspected false physiological state segments. Based on the physiological parameters, the remaining detection results that are determined to be physiological states are verified, and physiological state data segments that meet the preset false detection conditions are marked as suspected false detection physiological state segments. Determine the number of all false positives labeled as suspected physiological states and the sum of the probabilities of all false positives labeled as suspected physiological states being identified as physiological states, and calculate the self-evaluation adjustment coefficient.

7. The model parameter update method as described in claim 6, characterized in that, The step of updating the model parameters of the preset physiological state detection model based on the self-assessment adjustment coefficient, feedback confidence level, and historical feedback information coefficient includes: If the triggering condition is triggered autonomously by the model, then all labels marked as suspected false physiological states will be modified to non-physiological states, forming a second physiological state detection result; Based on the second physiological state detection result, obtain the updated physiological state feature array and corresponding labels; Obtain the historical physiological state feature array and corresponding labels, and combine it with the updated physiological state feature array and corresponding labels to update the dataset as parameters; The preset physiological state detection model is iteratively trained based on the updated dataset of the parameters to obtain the updated preset physiological state detection model.

8. The model parameter update method as described in claim 7, characterized in that, After updating the model parameters of the preset physiological state detection model based on the self-assessment adjustment coefficient, feedback confidence level, and historical feedback information coefficient, the method further includes: The dataset is updated based on the parameters, and the updated preset physiological state detection model is evaluated to obtain the number of suspected false physiological state segments. If the number of suspected false physiological state segments is less than a preset threshold, the historical physiological state feature dataset is updated. If the number of suspected false physiological state segments is not less than the preset threshold, the model parameter update step size is adjusted. If the number of times the model parameter update step size is adjusted is less than a preset threshold, the model is retrained based on the updated parameter dataset, and the number of suspected physiological state false detection segments in the updated model is evaluated. If the number of times the model parameter update step size is adjusted is not less than the preset threshold, the update fails and the original model parameters are retained.

9. A model parameter update system, characterized in that, The system includes: On the device side, features are extracted from the acquired physiological data of the target object, and the extracted features are input into a preset physiological state detection model for prediction to obtain physiological state detection results. The client is used to display the physiological state detection results to the user. If the user has feedback on the physiological state detection results, the user can input feedback information. The server-side component, when a preset trigger condition for updating the physiological state detection model parameters is met, acquires the physiological state detection results and physiological parameters for a target object within a target time period. The physiological state detection results include whether snoring is detected, and the physiological parameters include raw or basic data on the physiological function or activity characteristics of the target object in a bed-bound or sleep-like state. The trigger conditions include user feedback triggering and model-driven triggering. The server evaluates the physiological state detection results based on the physiological parameters, generating a self-evaluation adjustment coefficient. If the trigger condition is user feedback triggering, it performs logical consistency and reliability assessments based on the physiological state detection results, the physiological parameters, and user feedback information. Finally, it evaluates the self-evaluation adjustment coefficient, feedback confidence level, and historical data. The feedback information coefficient updates the model parameters of the preset physiological state detection model. The historical feedback information coefficient is determined by the trend and rate of change of historical feedback detection deviation levels. The self-assessment adjustment coefficient is determined based on the sum of probabilities that all suspected physiological state false detection segments are identified as the target physiological state and the number of suspected physiological state false detection segments. The suspected physiological state false detection segments are determined by the duration of the target physiological state, combined with the bed-leaning signal from the pressure sensor and the activity status of the motion sensor. The feedback confidence level is determined based on the number of suspected physiological state false detection segments reported by the user and the number of occurrences of actual physiological state events. The number of suspected physiological state false detection segments reported by the user is jointly determined by the physiological state detection results and user feedback information. The target physiological state refers to the snoring state.

10. The model parameter update system as described in claim 9, characterized in that, The device includes: The signal acquisition module is used to collect physiological data of the target object; The signal quality judgment module is used to filter the collected physiological data to obtain target physiological data that meets the quality requirements. The feature extraction module is used to extract features from the target physiological data to obtain physiological features; The physiological state detection module is pre-configured with the preset physiological state detection model, which is used to predict the physiological characteristics and obtain the physiological state detection results.

Images