A sleep multi-parameter fusion modeling method

Abstract

The application discloses a sleep multi-parameter fusion modeling method based on sleep, relates to the technical field of sleep data processing, and is used for solving the problem of poor data fusion credibility; the sleep parameters of a wristband, a mattress and environment equipment are mapped to a unified time reference through time unification, adaptive time slices are generated based on a change indicator, pollution segments caused by missing data and link jitter are isolated, individualized parameter state descriptions are further mapped from the trend, fluctuation and periodic characteristics in each slice, parameter state units are constructed, co-occurrence, precedence and repulsion relationships are extracted based on a state association graph, state consistency constraints are generated by combining deep learning sequence modeling, constraint propagation and consistency determination are performed on candidate fusion states, a continuous and interpretable fusion state result is output, a prediction window residual anomaly detection backtracking correction is introduced, and the constraints and the model are adaptively updated according to updated sample entries, so that the stability of the fusion result is improved.

Description

Technical Field

[0001] This invention relates to the field of sleep data processing technology, and more specifically, to a method for multi-parameter fusion modeling of sleep. Background Technology

[0002] In home sleep health management scenarios, sleep monitoring based on wristband, mattress sensors and indoor environmental devices has been widely used. Such solutions typically use mobile terminals or the cloud to aggregate and process parameters such as heart rate, body movement, respiratory rhythm, temperature, humidity and noise, and infer sleep stages and risk warnings through fixed time window slicing, threshold rules or conventional sequence models to support long-term sleep assessment and personalized recommendations. Its technical implementation involves multi-source time series data access, synchronization, feature extraction and state modeling.

[0003] However, inconsistent clock references between different devices, wireless link upload delays, and batch reporting mechanisms can introduce time misalignments. Missing measurements and noise can be amplified within a fixed slice, leading to misjudgments of parameter order and false triggering of state abrupt changes. At the same time, traditional fusion often relies on fixed thresholds or correlation assumptions from a single perspective, lacking constraints on the rationality of parameter combinations and a traceable judgment mechanism. This makes it difficult for the fusion results to remain continuous and stable when there is interference, missing data, or individual differences, thereby weakening the interpretability of sleep state characterization and the reliability of its use in health management decisions. Summary of the Invention

[0004] To overcome the aforementioned shortcomings of the existing technology, the following solutions are proposed to address the problem of poor data fusion reliability in the aforementioned background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for sleep multi-parameter fusion modeling includes the following steps:

[0007] S1. Obtain sleep parameters generated during sleep, perform time-consistent processing on the sleep parameters, dynamically adjust the length of the time slice, divide the continuous sleep process into multiple adaptive time slices, and form a corresponding time series set of sleep parameters.

[0008] S2. Within each time slice, the sleep parameters are processed to represent their state. The change characteristics of the sleep parameters within the corresponding time slice are mapped to personalized parameter state descriptions, and a parameter state unit is constructed to uniformly express the state of the sleep parameters.

[0009] S3. Based on the parameter state units within the same time slice, analyze the collaborative relationship between different parameter states, generate a state consistency constraint model to limit the rationality of parameter state combinations, and use a deep learning model to model long-term and short-term dependencies.

[0010] S4. On continuous time slices, a state consistency constraint model is introduced to perform fusion modeling of parameter state units. A fusion state result is generated to characterize the evolution process of sleep state through consistency judgment. The fusion result is corrected based on the anomaly detection mechanism.

[0011] S5. Update the state consistency constraint model based on the fusion state results. Through an adaptive learning mechanism, optimize the model parameters and state consistency constraints in real time so that the fusion modeling process can adapt to the changes in parameter state correlations in different sleep processes.

[0012] Furthermore, sleep parameters are subjected to time-consistency processing, including:

[0013] Determine the target time base;

[0014] Calculate the timestamp offset for the data stream of each sleep parameter and perform offset correction;

[0015] Determine the target sampling time point set, and perform resampling mapping on the data stream of each sleep parameter on the target sampling time point set to generate a parameter sequence aligned with the time points;

[0016] Missing segments in the resampling map are generated with missing test markers and written into the sleep parameter time series set;

[0017] Dynamically adjusting the time slice length includes: calculating change indicators based on the parameter sequence of aligned time points, determining slice boundaries based on the change indicators, and generating an adaptive time slice sequence.

[0018] Furthermore, within each time slice, sleep parameters are subjected to state characterization processing, including:

[0019] Denoising and amplitude normalization are performed on the parameter sequence within the time slice;

[0020] Extract trend features, fluctuation features, and periodic features used to characterize changes to form state feature entries;

[0021] Based on the state feature entries, determine the parameter state category from the predefined state set, and output a parameter state description containing the parameter identifier, parameter state category, and state feature entries.

[0022] Furthermore, a parameter state unit for uniformly expressing sleep parameter states is constructed, including:

[0023] Write the time slice identifier, parameter identifier, parameter status description, change indicator, and confidence flag into the same status unit;

[0024] Personalized parameter status descriptions are obtained through individual baseline generation, which includes establishing an individual parameter distribution baseline based on a historical sleep parameter time series set.

[0025] The boundary rules of the state set are determined based on the baseline of individual parameter distribution, and the parameter state category is determined based on the boundary rules.

[0026] Furthermore, the cooperative relationships between different parameter states are analyzed, including:

[0027] Within the same time slice, a state association diagram is constructed based on the parameter identifier and parameter state category in the parameter state unit;

[0028] Based on the state association graph, relationship feature entries of co-occurrence, sequence and exclusion relationships are extracted;

[0029] The relation feature entries on continuous time slices are combined into a relation sequence and input into a deep learning sequence modeling network. The output is a relation representation vector used to characterize the long-term and short-term dependencies of parameter states.

[0030] Furthermore, a state consistency constraint model is generated to limit the rationality of parameter state combinations, including:

[0031] A set of constraint rule entries is generated based on the relation representation vector. Each constraint rule entry contains a triggering condition and a constraint conclusion.

[0032] Triggering conditions are used to limit the parameter state combination mode within a time slice or the state transition mode between adjacent time slices, and constraint conclusions are used to indicate allowed combinations, prohibited combinations, and preferred combinations.

[0033] Write the set of constraint rule entries into the state consistency constraint model and form a searchable constraint index.

[0034] Furthermore, the parameter state units are fused and modeled, including:

[0035] Construct a candidate fusion state set on continuous time slices and establish a mapping relationship between the candidate fusion state set and the parameter state unit;

[0036] A state consistency constraint model is introduced to perform constraint propagation on the candidate fusion state set, generating a set of feasible fusion paths that satisfy the constraint rule entries;

[0037] Perform a consistency determination on the set of feasible fusion paths, select the fusion path with the fewest number of violation rule entries and the fewest state transition conflicts, and output the fusion state result corresponding to the fusion path.

[0038] Furthermore, the fusion state results are corrected based on the anomaly detection mechanism, including:

[0039] A fusion state prediction window is established based on the fusion path. The prediction window is used to generate the predicted fusion state within the time slice.

[0040] Calculate the consistency residual between the predicted fusion state and the fusion state result, and generate anomaly markers based on the consistency residual;

[0041] When the anomaly marker meets the triggering condition, backtrack the time slice window corresponding to the anomaly marker and re-execute the candidate fusion state construction and constraint propagation, output the corrected fusion state result and record the correction window identifier.

[0042] Furthermore, the state consistency constraint model is updated, including:

[0043] An updated sample entry is generated based on the corrected fusion state result. The updated sample entry includes time slice identifier, parameter state unit, fusion state result and anomaly marker.

[0044] The number of times the constraint rule entries are satisfied and the number of times they are violated are statistically analyzed based on the updated sample entries, and the trigger condition boundaries of the constraint rule entries are adaptively adjusted.

[0045] Incremental training is performed on the deep learning sequence modeling network based on updated sample entries, so that the relation representation vector adapts to the changes in the parameter state association relationship.

[0046] The technical effects and advantages of the present invention based on a multi-parameter sleep fusion modeling method are as follows:

[0047] This invention achieves comparable and traceable representation of sleep parameters from multiple devices on a unified time reference in home sleep health management scenarios by constructing a closed-loop fusion modeling process that includes sleep parameter time consistency, adaptive time slicing, parameter state unit construction, state consistency constraint modeling, and fusion result updating. It effectively suppresses the sequential misalignment and missing measurement contamination caused by wristband batch reporting, wireless link jitter, and short-term mattress absence, and avoids misjudging device delays or absences as sudden changes in the actual sleep state.

[0048] Building upon this foundation, the model maps the trends, fluctuations, and cyclical features within each time slice to individualized parametric state descriptions, forming parametric state units. A state association graph is introduced to extract co-occurrence, sequence, and exclusion relationships. Deep learning sequence modeling is then combined to obtain long- and short-term dependency relationship representation vectors, thereby generating a searchable state consistency constraint model. Constraint propagation and consistency judgment are performed on candidate fusion states, resulting in continuous and interpretable fusion state results. Simultaneously, a prediction window residual anomaly detection is introduced to backtrack and correct inconsistent segments and record correction window identifiers. Finally, based on updated sample entries, the constraint rule trigger boundaries and sequence networks are adaptively updated, enabling the model to continuously calibrate according to individual differences, wearing tightness, and environmental disturbances. This reduces misjudgments and state jumps, improving the stability, robustness, and interpretability of sleep state characterization, providing a more reliable data foundation for subsequent health assessments and intervention decisions. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating a sleep multi-parameter fusion modeling method according to the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] In order to achieve the above objectives, Figure 1 A schematic diagram of a sleep multi-parameter fusion modeling method based on the present invention is given, which specifically includes the following steps;

[0052] S1. Obtain sleep parameters generated during sleep, perform time-consistent processing on the sleep parameters, dynamically adjust the length of the time slice, divide the continuous sleep process into multiple adaptive time slices, and form a corresponding time series set of sleep parameters.

[0053] S2. Within each time slice, the sleep parameters are processed to represent their state. The change characteristics of the sleep parameters within the corresponding time slice are mapped to personalized parameter state descriptions, and a parameter state unit is constructed to uniformly express the state of the sleep parameters.

[0054] S3. Based on the parameter state units within the same time slice, analyze the collaborative relationship between different parameter states, generate a state consistency constraint model to limit the rationality of parameter state combinations, and use a deep learning model to model long-term and short-term dependencies.

[0055] S4. On continuous time slices, a state consistency constraint model is introduced to perform fusion modeling of parameter state units. A fusion state result is generated to characterize the evolution process of sleep state through consistency judgment. The fusion result is corrected based on the anomaly detection mechanism.

[0056] S5. Update the state consistency constraint model based on the fusion state results. Through an adaptive learning mechanism, optimize the model parameters and state consistency constraints in real time so that the fusion modeling process can adapt to the changes in parameter state correlations in different sleep processes.

[0057] S1. Obtain sleep parameters generated during sleep, perform time-consistency processing on these parameters, dynamically adjust the time slice length, divide the continuous sleep process into multiple adaptive time slices, and form a corresponding time series set of sleep parameters. Specifically, the implementation is as follows:

[0058] In a home sleep health management scenario, users simultaneously collect heart rate and body movement parameters using a wristband device, respiratory rhythm parameters using a mattress sensor, and temperature, humidity, and noise parameters using an indoor environment device during nighttime sleep. The mobile terminal acts as a data aggregation point, receiving data streams uploaded from each device. The data processing end is used to perform time consistency, slicing, state representation, constraint modeling, fusion judgment, and update training. The data processing end can be implemented by the local processing module of the mobile terminal or by a cloud processing server. The mobile terminal is used to complete data aggregation and result display. The two transmit the time series set of sleep parameters and the fusion state results through network communication.

[0059] Collect sleep parameters generated during sleep and perform time-consistent processing on these parameters, specifically including:

[0060] The target time base is determined, and the reference time axis of the data processing end is used as a unified reference. The timestamp offset is calculated for the data stream of each sleep parameter and the offset correction is performed. The timestamp offset is determined by the correspondence between the source timestamp reported by the device and the receiving time of the mobile terminal. During the correction, the offset is applied to the timestamp field of the device's data stream so that each data stream is mapped to the same target time base.

[0061] In one embodiment, the timestamp offset and correction specifically includes:

[0062] Timestamp offset of the j-th sleep parameter data stream The correspondence between the source timestamp and the receiving time is determined, preferably using robust statistics: ,in, This represents the source timestamp of the k-th sample in the j-th data stream. This indicates the time when the corresponding sample arrives at the mobile terminal for reception. It is a median function;

[0063] The correspondence of the kth sample is established by the sample number or message identifier. When there is no message identifier, the nearest arrival time pairing is adopted. The message with the closest reception time is selected as the corresponding sample within the preset pairing window. Messages outside the pairing window are not included in the median estimation and are recorded as outlier samples.

[0064] The correction timestamp is obtained after offset correction. : Through the above corrections, the sleep parameter data streams are mapped to the same target time reference.

[0065] After offset correction is completed, a target sampling time point set is determined. This target sampling time point set is a discrete sampling time sequence on a unified time axis, used to carry the alignment results of parameters from different sources. Resampling mapping is then performed on the data streams of each sleep parameter on the target sampling time point set to generate the parameter sequence of the aligned time points.

[0066] For continuous parameters, piecewise interpolation mapping is used to obtain the aligned sequence. Linear interpolation is preferred for piecewise interpolation mapping.

[0067] For target sampling times that fall between two adjacent source times, aligned samples are obtained by linear interpolation of the sample values ​​at both ends. When the interval between adjacent source times exceeds the maximum interpolation span, the target sampling time is recorded as missing and a missing measurement marker is generated.

[0068] For event-type or sparse parameters, the nearest neighbor mapping or window aggregation mapping is used to obtain the alignment sequence. The nearest neighbor mapping uses the source sample closest to the target sampling time as the alignment value. The window aggregation mapping constructs an aggregation window with the target sampling time as the center and writes the event count, maximum amplitude or duration within the window into the alignment sample. When there are no events in the window, null values ​​are written and a missing test mark is set.

[0069] If the original data stream cannot support an effective mapping to the target sampling time point within a certain time interval, the segment is identified as a missing segment. Missing test markers are generated for the missing segments in the resampling mapping, and the missing test markers and the alignment sequence of the corresponding time points are written into the sleep parameter time series set.

[0070] For example, during the same night's sleep, the wristband device collects heart rate and body movement data, and the mattress sensor device collects respiratory rhythm data. Because the mattress device uploads data via a wireless link, when the user turns over and presses against the edge of the mattress, briefly blocking the signal, the respiratory rhythm data may not be reported for a period of time.

[0071] Meanwhile, due to the local caching and batch uploading mechanism of the wristband device, the source timestamp reported by the device has a stable offset relative to the time of reception by the mobile terminal. If time consistency processing is not performed, the physical movement event will appear to occur after the heart rate change on the mobile terminal timeline, which can easily be misjudged as the heart rate change caused by physical movement.

[0072] Therefore, after performing timestamp offset correction and completing alignment, the mobile terminal observes body movement and heart rate changes occurring within the same time range under a unified time reference. This avoids mistaking device reporting delays for physiological causal relationships. When the mattress breathing rhythm data cannot be mapped to the target sampling time point set during the turning period, the server identifies this segment as a missing segment, writes a missing test marker, and saves it together with the alignment sequence of other parameters into the sleep parameter time series set. Subsequently, during dynamic slicing, the server uses the start and end positions of the missing segment as the slice boundary, making the missing segment an independent adaptive time slice, and synchronously writes the missing test impact marker into the adaptive time slice sequence. Through this processing, it avoids mistaking abnormal fluctuations caused by breathing absence as real sleep state changes, and also avoids data from missing segments polluting the state modeling results of adjacent normal slices, thereby improving the usability and interpretability of the fusion modeling input.

[0073] The server dynamically adjusts the time slice length, dividing the continuous sleep process into multiple adaptive time slices. The dynamic adjustment process includes:

[0074] The server calculates a change indicator based on the parameter sequence at aligned time points. The change indicator is used to describe the overall change intensity and consistency of sleep parameters within adjacent time ranges.

[0075] In this embodiment, the change indication quantity consists of two parts:

[0076] The first is the intensity of change, which is formed by creating a change sequence based on the absolute change of each sleep parameter at adjacent sampling points, and then taking the median of each parameter change sequence to obtain the intensity. The second is the consistency of change, which is obtained by statistically analyzing the proportion of parameters with the same change direction within the same time range. When both the intensity and consistency are high, it is judged as a candidate for the boundary of a strong change segment. When the intensity is low and the consistency is stable, the slice span is extended.

[0077] When calculating change indicators, the server performs masking or de-confidence processing on segments marked with missing data to prevent missing data from dominating slicing decisions.

[0078] The server determines the slice boundaries based on the change indicator: when the change indicator is at a high level and continues to change, the slice boundaries are set to be denser to obtain finer-grained slices.

[0079] When the change indicator is at a low level and stable, extend the slice span to obtain a more stable slice.

[0080] When missing markers appear in clusters within a local time period, the edges of the missing segments are set as slice boundaries to isolate incomplete fragments and reduce interference with subsequent modeling. This generates an adaptive time slice sequence.

[0081] The output of step S1 is a time series set of sleep parameters and an adaptive time slice sequence. The time series set of sleep parameters includes at least: target time reference identifier, target sampling time point set identifier, alignment sequence of each parameter, and missing test marker.

[0082] An adaptive time-slice sequence should include at least: slice identifier, slice start and end times, a summary of changes in the slice boundary generation basis, and a marker for missing data impact.

[0083] S2. Within each time slice, the sleep parameters are processed to represent their state. The changes in sleep parameters within the corresponding time slice are mapped to personalized parameter state descriptions, and a parameter state unit for uniformly expressing the sleep parameter state is constructed. The specific implementation is as follows:

[0084] For a given time slice, first extract the parameter sequence that falls within the start and end time range of the time slice from the sleep parameter time series set, and then perform state representation processing.

[0085] In the state characterization process, the parameter sequences within the time slice are subjected to denoising and amplitude normalization: for continuous sequences such as heart rate and respiratory rhythm, smoothing filtering or outlier suppression is used to weaken instantaneous spike interference; for sparse change sequences such as body movement, window aggregation is used to suppress occasional jitter.

[0086] Environmental sequences such as temperature, humidity, and noise are smoothed by preserving the gradual change trend.

[0087] After denoising is completed, the server performs amplitude normalization on different parameter sequences under the same normalization rule, so that parameter sequences with different dimensions can be processed by the same standard in subsequent feature extraction and category determination.

[0088] For sample points with missing test markers, the missing test markers are retained and masked or down-confidenced during normalization to avoid missing values ​​affecting feature calculation.

[0089] Subsequently, the server extracts trend features, fluctuation features, and periodic features to characterize the changes, forming state feature entries. Trend features are used to describe the overall upward, downward, or stable direction of the parameter within the time slice; fluctuation features are used to describe the intensity and instability of the parameter's fluctuations within the slice; and periodic features are used to describe rhythmic information with repetitive patterns, such as respiratory rhythm and body movement intervals.

[0090] Trend characteristics are determined by taking the first difference of the parameter sequence within the slice and taking the sign of the median. Fluctuation characteristics are characterized by calculating the absolute deviation of the median of the parameter sequence within the slice to represent the intensity of fluctuations. Periodic characteristics are characterized by whether the autocorrelation peak or the main peak in the frequency domain appears stably.

[0091] The above features are written into the state feature entries as structured fields, and the state feature entries are associated with the parameter identifiers to form a feature set for this time slice.

[0092] After generating the state feature entries, the parameter state category is determined from the predefined state set based on the state feature entries, and a parameter state description containing the parameter identifier, parameter state category and state feature entries is output.

[0093] A predefined set of states is used to provide a set of discriminative state category labels for each type of parameter. Instead of directly using a fixed threshold, individual baseline generation is introduced to obtain personalized parameter state descriptions.

[0094] Individual baseline generation includes:

[0095] The server extracts multiple sleep records of the same user from the historical sleep parameter time series set to establish an individual parameter distribution baseline. The individual parameter distribution baseline is used to characterize the common distribution patterns and fluctuation ranges of each parameter of the user in different sleep stages. Based on the individual parameter distribution baseline, the server determines the boundary rules of the state set. The boundary rules are used to map state feature entries to state category labels. For example, a stable trend with low fluctuation is mapped to the stable class, an upward trend with high fluctuation is mapped to the active class, and a clear periodicity with moderate fluctuation is mapped to the rhythm class.

[0096] The server determines the parameter state category based on boundary rules, obtaining a personalized parameter state description, making the same state category comparable for different individuals and reflecting individual differences.

[0097] It should be noted that the boundary rules are generated based on the quantile intervals of the individual parameter distribution baseline. Trend features falling into the stable interval and fluctuation features falling into the low fluctuation interval are classified as stable; trend features falling into the rising interval and fluctuation features falling into the high fluctuation interval are classified as active; periodic features satisfying the stable main peak and fluctuation features falling into the medium fluctuation interval are classified as rhythmic; when missing markers are concentrated or the proportion of effective samples is insufficient, the parameter state category remains uncertain and the confidence level is reduced.

[0098] After obtaining the parameter status description for each parameter, the server constructs a parameter status unit to uniformly express the sleep parameter status. The parameter status unit is organized by time slices, and the time slice identifier, parameter identifier, parameter status description, change indicator and confidence flag are written into the same status unit.

[0099] The change indicator is taken from the change indicator used in step S1 to determine the slice boundary and is associated with the parameter state description within the slice to distinguish between slices with strong and weak changes. The confidence marker is determined jointly by the missing data marker, the denoising intensity, and the normalized effective sample ratio.

[0100] When missing test markers appear in clusters or there are insufficient valid samples, the server lowers the confidence marker; when the slice data is complete and feature extraction is stable, the server raises the confidence marker.

[0101] In one embodiment, the proportion of effective samples within a time slice is denoted as: ,in, This represents the total number of sampling points within the slice. The number of sampling points marked as 1 for missing data is given. The confidence level is preferably determined by a combination of the effective sample ratio and the denoising intensity. When the value falls below a preset threshold, the confidence flag is directly reduced, which serves as a pruning strategy for subsequent consistency determination and anomaly backtracking.

[0102] The parameter state units corresponding to each parameter within the same time slice are collected to form the state unit set of that time slice, and output in chronological order as the input for the subsequent step S3, so as to analyze the cooperative relationship between different parameter states within the same time slice and generate a state consistency constraint model.

[0103] S3. Based on parameter state units within the same time slice, analyze the collaborative relationships between different parameter states, generate a state consistency constraint model to limit the rationality of parameter state combinations, and use a deep learning model to model long- and short-term dependencies. The specific implementation is as follows:

[0104] The server takes a set of state units for a specific time slice as input and constructs a state association graph based on the parameter identifiers and parameter state categories within the parameter state units. The state association graph uses nodes corresponding to parameter identifiers to represent different sleep parameters, and the state category of each node as its attribute.

[0105] The server establishes association edges between node pairs within the same time slice and writes the co-occurrence strength, mutual exclusion flag, and sequential pointing flag for the association edges within that slice. The co-occurrence strength is used to characterize the degree to which two parameter state categories appear simultaneously within the slice. The mutual exclusion flag is used to characterize the conflict relationship that two state categories should not be simultaneously established within the same slice. The sequential pointing flag is used to characterize the first-occurrence and second-occurrence relationship within the slice or near the slice boundary. This forms a state association graph oriented towards a single time slice and generates a corresponding graph sequence as the time slice progresses.

[0106] After constructing the state association graph, the server extracts relationship feature entries based on the state association graph, including co-occurrence, sequence, and exclusion relationships. Each relationship feature entry describes a pair of parameter identifiers and their state category combinations using structured fields, and includes co-occurrence scores, sequence direction, and exclusion markers. For example, if the state category corresponding to body movement is highly active in a certain slice, while the state category corresponding to respiratory rhythm is unstable in the same slice, the server records the co-occurrence scores of body movement and respiration in the relationship feature entries of that slice.

[0107] When the body movement state category changes first and the heart rate state category changes subsequently in adjacent slices, the server records the order of the relationship in the relational feature entries. When a certain parameter state category and another parameter state category exhibit a conflicting pattern in the same slice, the server records an exclusion flag in the relational feature entries and reduces the confidence of the combination. The server repeats the above extraction process for each time slice to obtain a sequence of relational feature entries that change with time slices.

[0108] Subsequently, the server assembles the relation feature entries from consecutive time slices into a relation sequence and inputs it into a deep learning sequence modeling network. The output is a relation representation vector characterizing the long-term and short-term dependencies of the parameter states. The deep learning sequence modeling network takes the relation sequence as input, learns short-term dependencies to depict rapid linkages between adjacent slices, and learns long-term dependencies to depict slowly varying relationships across multiple slices. The server writes the relation representation vector output by the network into the relation representation cache of the current sleep process and establishes an association with the corresponding time range and relation feature entry sequence, so that the corresponding relation source can be traced back when generating constraint rules later.

[0109] Deep learning sequence modeling networks can be implemented using gated recurrent unit networks or self-attention encoders. The input is a sequence of relation feature entries ordered by time slices, and the output is a relation representation vector. The training objective is to keep the relation representations of adjacent slices temporally consistent and to make the relation representations corresponding to the violation rule entries separable in the feature space.

[0110] After obtaining the relation representation vector, the server generates a state consistency constraint model to limit the rationality of parameter state combinations. The generation process includes: generating a set of constraint rule entries based on the relation representation vector. Each constraint rule entry contains a triggering condition and a constraint conclusion. The triggering condition is used to limit the parameter state combination pattern within a time slice or the state transition pattern between adjacent time slices. The triggering condition is composed of parameter identifier, parameter state category, slice position relationship, and confidence flag condition. The constraint conclusion is used to indicate allowed combinations, prohibited combinations, and preferred combinations. In the case of prohibited combinations, a conflict attribution type is given, and in the case of preferred combinations, a priority flag is given.

[0111] The set of constraint rule entries is written into the state consistency constraint model to form a searchable constraint index. The constraint index uses parameter identifier pairs, state category pairs, and slice transition patterns as search keys.

[0112] S4. On continuous time slices, a state consistency constraint model is introduced to perform fusion modeling of parameter state units. A fusion state result representing the evolution process of sleep state is generated through consistency determination. The fusion result is corrected based on an anomaly detection mechanism. The specific implementation is as follows:

[0113] The server first constructs a candidate fusion state set on continuous time slices and establishes a mapping relationship between the candidate fusion state set and the parameter state unit. The candidate fusion state set is generated based on the fusion state category set, which is used to uniformly express the overall sleep state category of the time slice.

[0114] For each time slice, the server generates candidate fusion state entries based on the set of state units within that slice. Each candidate fusion state entry contains at least a fusion state category, a summary of parameter state combinations that support that category, and a confidence tag summary.

[0115] It should be noted that the fusion state category set includes at least a portion of wakefulness, light sleep, deep sleep, and REM sleep. The fusion state results are output in units of time slices, along with the corresponding category and confidence marker.

[0116] The mapping relationship is used to associate candidate fusion state entries with the set of parameter state units of the slice, so that subsequent constraint propagation can locate which parameter states support the candidate fusion state and which parameter states constrain it.

[0117] Subsequently, the server introduces a state consistency constraint model to perform constraint propagation on the candidate fusion state set, generating a set of feasible fusion paths that satisfy the constraint rule entries, specifically including:

[0118] The server uses time slice sequences as the order and connects candidate fusion state entries into path candidates according to the adjacency relationship of slices. When expanding path candidates each time, the server retrieves the constraint rule entries corresponding to the parameter state combination mode in the current slice through the constraint index, and retrieves the constraint rule entries corresponding to the state transition mode between adjacent slices. Path candidates that do not meet the prohibited combination or prohibited transition are pruned, and path candidates that meet the allowed combination and preferred combination are retained and sorted. This process is an implementation method of constraint propagation and pruning in the candidate space, which can make feasible paths converge to the subspace that satisfies the consistency constraints more quickly.

[0119] After obtaining the set of feasible fusion paths, the server performs a consistency determination on the set of feasible fusion paths, selects the fusion path with the fewest number of violation rule entries and the fewest state transition conflicts, and outputs the fusion state result corresponding to the fusion path.

[0120] In the consistency determination, the server counts the number of triggered but unmet constraint rule entries for each feasible fusion path, and counts the number of transition conflicts where the fusion state categories are discontinuous or mutually exclusive between adjacent slices. When multiple paths satisfy the same number of default rule entries and the same number of conflicts, the server prioritizes the path containing more priority combination conclusion entries. The final output fusion state results are given in time slice order, with each time slice corresponding to a fusion state category and its confidence flag, and is associated with the parameter state unit set of that slice to form a traceable fusion result chain.

[0121] After outputting the fusion status results, the server initiates an anomaly detection mechanism to correct them. The server establishes a fusion status prediction window based on the fusion path. This prediction window generates the predicted fusion status within a time slice, including:

[0122] The prediction window selects the fusion state results of adjacent slices as the context with the current slice as the center, and combines them with the constraint rule entries in the constraint model corresponding to the window to obtain the predicted fusion state category of the current slice.

[0123] The server calculates the consistency residual between the predicted fusion state and the fusion state result, and generates anomaly markers based on the consistency residuals: when the predicted fusion state category is inconsistent with the fusion state result category and the inconsistency is persistent or abrupt within the window, the anomaly marker is set to valid. This type of residual anomaly detection based on the deviation between prediction and reality is a common implementation approach for time series anomaly detection. The prediction window is used to provide local temporal context to enhance the stability of the judgment.

[0124] When the anomaly marker meets the triggering condition, the server backtracks the time slice window corresponding to the anomaly marker and re-executes the candidate fusion state construction and constraint propagation, outputs the corrected fusion state result and records the corrected window identifier. In this embodiment, the triggering condition is jointly limited by the validity of the anomaly marker, the persistence of the anomaly within the window and the instability of the candidate path.

[0125] After backtracking, the server regenerates the candidate fusion state entries in the window and increases the priority of the constraint strength of prohibited combination and prohibited transfer entries during the constraint propagation phase, so that the correction results return to the fusion path consistent with the parameter state coordination relationship. The correction window identifier and the fusion state results before and after the correction are written into the process record together.

[0126] After outputting the fusion state result and completing the anomaly detection and correction in step S4, the mobile terminal enters step S5 to update the state consistency constraint model. This allows the model to gradually absorb changes in the parameter state correlation caused by the same user under different nights, different wearing tightness, and different environmental disturbances, thus avoiding rule distortion and fusion path deviation during long-term use.

[0127] For example, during the same nighttime sleep cycle, the server outputs a merged path across consecutive time slices. A typical short-term anomaly window appears during this period:

[0128] When a user briefly wakes up in the middle of the night to adjust their sleeping position, the wristband movement shows significant activation within the window, and the heart rate also shows an upward trend. However, the mattress breathing rhythm shows missing markers in the first half of the window due to poor contact, resulting in a decrease in the confidence marker of the breathing parameter status unit within the window.

[0129] According to the state consistency constraint model established in step S3, when body movement is activated and accompanied by an increase in heart rate, the fusion path should usually prioritize the fusion state category consistent with the combination, and require the state transition between adjacent slices to remain continuous; however, within this abnormal window, due to insufficient confidence in the respiratory state unit, two competing path candidates appear in the candidate fusion state set: one path is more consistent with the synergistic relationship between body movement and heart rate, while the other path has a transition conflict in the consistency judgment due to respiratory uncertainty caused by missing measurements;

[0130] After the server establishes a prediction window based on the fusion path, the prediction window combines the fusion state results of adjacent slices with the constraint rule entries to give the predicted fusion state that the window should maintain continuity and be consistent with the combination of body movement and heart rate.

[0131] When the actual fusion state is inconsistent with the predicted fusion state, and the inconsistency exhibits abrupt changes within the window, the server generates an anomaly marker and triggers a backtracking.

[0132] After backtracking, the server only reconstructs the candidate fusion state set for the abnormal window and re-executes constraint propagation: during propagation, prohibition transitions and prohibition combinations rules are pruned more strictly, while priority is increased for priority combination rules that match the coordination relationship between body movement and heart rate. Finally, the corrected fusion state result is output, and the correction window identifier is recorded. This corrected result is written into the updated sample entries. When the pattern of turning over repeatedly occurs in subsequent sleep cycles, resulting in short-term absence of mattress measurements but stable wristband signals, the server can adaptively adjust the trigger condition boundaries of relevant constraint rule entries by updating sample entries. This allows the model to gradually learn not to easily trigger unreasonable state transitions under conditions of respiratory measurement absence and decreased confidence, thereby reducing misjudgments and improving the temporal continuity of the fusion state results.

[0133] S5. Update the state consistency constraint model based on the fusion state results. Through an adaptive learning mechanism, optimize the model parameters and state consistency constraints in real time to adapt the fusion modeling process to changes in parameter state correlations during different sleep processes. Specifically, the implementation is as follows:

[0134] The server generates updated sample entries based on the corrected fusion state results. The generation process establishes entry records one by one in time slice order, specifically including:

[0135] Each updated sample entry includes at least a time slice identifier, the set of parameter state units corresponding to the slice, the fusion state result corresponding to the slice, the anomaly flag corresponding to the slice, and the correction window identifier's attribution information for the slice. If a time slice is within the correction window, the server sets the anomaly flag to valid in the updated sample entry and writes the correction window identifier to distinguish the source difference between the original fusion output and the corrected fusion output. If a time slice does not trigger correction, the anomaly flag is set to invalid to indicate that the fusion result of the slice can directly participate in the rule satisfaction statistics.

[0136] Subsequently, the server updates the sample entries to count the number of times the constraint rule entries are satisfied and the number of times they are violated, and adaptively adjusts the trigger condition boundaries of the constraint rule entries. Specifically, the server maintains an event counter and a time decay window index for each constraint rule entry in the state consistency constraint model.

[0137] When an updated sample entry arrives, the server uses the constraint index to retrieve the constraint rule entries triggered within that time slice and the constraint rule entries triggered between adjacent time slices, checks whether the triggering conditions are met for each entry, and determines whether the constraint conclusion is satisfied by the fusion path.

[0138] If the condition is met, the number of times it is met will be incremented; if it is violated, the number of times it is violated will be incremented. The location of the occurrence will be associated with the time slice identifier. The server will parameterize the trigger condition boundary into an updatable boundary field, such as the boundary of trend feature, the boundary of fluctuation feature, the boundary of cycle feature, the allowable span of sequential relationship, and the threshold of the strength of the exclusion relationship as part of the trigger condition.

[0139] When a rule accumulates a continuous number of violations during a period of continuous sleep, the server adaptively adjusts the boundary fields of that rule:

[0140] Extract state feature entries and relationship feature entries associated with the rule from the updated sample entries, calculate the distribution range of these entries within the recent window, and move the boundary field into the distribution range to reduce systemic false triggers caused by changes in individual state or equipment conditions.

[0141] Boundary field updates are achieved by re-estimating the quantile intervals of the features corresponding to the recently updated sample entries. When a rule is satisfied a stable number of times and the number of violations is extremely low, the server keeps the boundary field unchanged and increases the priority label of the rule in the constraint propagation stage, making it easier for it to play a constraint role in subsequent fusion path pruning.

[0142] After completing the statistical analysis and boundary adaptive adjustment of the constraint rule entries, the server performs incremental training on the deep learning sequence modeling network based on the updated sample entries, so that the relation representation vector adapts to the changes in the parameter state association relationship.

[0143] Specifically, the server reconstructs the sequence of relational feature entries from the updated sample entries, including:

[0144] For each time slice, the corresponding state association graph and relation feature entries are traced back based on the parameter state unit set and the fusion state result, and a relation sequence is formed in chronological order. The server manages samples in the window that contain valid anomaly labels separately from samples that contain invalid anomaly labels. For samples that contain valid anomaly labels, the training order is adopted to first correct the relation representation and then use them to generate rules, so as to avoid the relation representation vector being incorrectly pulled by anomalies.

[0145] Incremental training adopts a mini-batch update method: each time, a sequence of relationships in a continuous time slice is used as input, and the network outputs a relationship representation vector. The consistency training objective is constructed with the satisfaction of the current set of constraint rule entries, so that the network can gradually learn new short-term linkage patterns and long-term dependency patterns. Since sleep sequence modeling itself has the characteristics of long-term and short-term dependencies, the sequence modeling network continuously absorbs new nighttime relationship sequences.

[0146] After each corrected fusion state result is formed, a closed-loop update is completed by updating sample entries, including rule statistics, boundary adaptive adjustment, and incremental training of the sequence network. This allows the state consistency constraint model and the deep learning sequence modeling network to gradually adapt to changes in user sleep behavior and collection conditions.

[0147] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0148] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0149] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0150] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0151] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for multi-parameter sleep fusion modeling, characterized in that: The specific steps include: S1. Obtain sleep parameters generated during sleep, perform time-consistent processing on the sleep parameters, dynamically adjust the length of the time slice, divide the continuous sleep process into multiple adaptive time slices, and form a corresponding time series set of sleep parameters. S2. Within each time slice, the sleep parameters are processed to represent their state. The change characteristics of the sleep parameters within the corresponding time slice are mapped to personalized parameter state descriptions, and a parameter state unit is constructed to uniformly express the state of the sleep parameters. S3. Based on the parameter state units within the same time slice, analyze the collaborative relationship between different parameter states, generate a state consistency constraint model to limit the rationality of parameter state combinations, and use a deep learning model to model long-term and short-term dependencies. S4. On continuous time slices, a state consistency constraint model is introduced to perform fusion modeling of parameter state units. A fusion state result is generated to characterize the evolution process of sleep state through consistency judgment. The fusion result is corrected based on the anomaly detection mechanism. S5. Update the state consistency constraint model based on the fusion state results. Through an adaptive learning mechanism, optimize the model parameters and state consistency constraints in real time so that the fusion modeling process can adapt to the changes in parameter state correlations in different sleep processes.

2. The sleep multi-parameter fusion modeling method according to claim 1, characterized in that: Time-consistent processing of sleep parameters includes: Determine the target time base; Calculate the timestamp offset for the data stream of each sleep parameter and perform offset correction; Determine the target sampling time point set, and perform resampling mapping on the data stream of each sleep parameter on the target sampling time point set to generate a parameter sequence aligned with the time points; Missing segments in the resampling map are generated with missing test markers and written into the sleep parameter time series set; Dynamically adjusting the time slice length includes: calculating change indicators based on the parameter sequence of aligned time points, determining slice boundaries based on the change indicators, and generating an adaptive time slice sequence.

3. The method for sleep multi-parameter fusion modeling according to claim 2, characterized in that: Within each time slice, sleep parameters are characterized by state processing, including: Denoising and amplitude normalization are performed on the parameter sequence within the time slice; Extract trend features, fluctuation features, and periodic features used to characterize changes to form state feature entries; Based on the state feature entries, determine the parameter state category from the predefined state set, and output a parameter state description containing the parameter identifier, parameter state category, and state feature entries.

4. The sleep multi-parameter fusion modeling method according to claim 3, characterized in that: Construct a parameter state unit for uniformly representing sleep parameter states, including: Write the time slice identifier, parameter identifier, parameter status description, change indicator, and confidence flag into the same status unit; Personalized parameter status descriptions are obtained through individual baseline generation, which includes establishing an individual parameter distribution baseline based on a historical sleep parameter time series set. The boundary rules of the state set are determined based on the baseline of individual parameter distribution, and the parameter state category is determined based on the boundary rules.

5. The sleep multi-parameter fusion modeling method according to claim 4, characterized in that: Analyze the cooperative relationships between different parameter states, including: Within the same time slice, a state association diagram is constructed based on the parameter identifier and parameter state category in the parameter state unit; Based on the state association graph, relationship feature entries of co-occurrence, sequence and exclusion relationships are extracted; The relation feature entries on continuous time slices are combined into a relation sequence and input into a deep learning sequence modeling network. The output is a relation representation vector used to characterize the long-term and short-term dependencies of parameter states.

6. The sleep multi-parameter fusion modeling method according to claim 5, characterized in that: Generate a state consistency constraint model to limit the rationality of parameter state combinations, including: A set of constraint rule entries is generated based on the relation representation vector. Each constraint rule entry contains a triggering condition and a constraint conclusion. Triggering conditions are used to limit the parameter state combination mode within a time slice or the state transition mode between adjacent time slices, and constraint conclusions are used to indicate allowed combinations, prohibited combinations, and preferred combinations. Write the set of constraint rule entries into the state consistency constraint model and form a searchable constraint index.

7. The sleep multi-parameter fusion modeling method according to claim 6, characterized in that: The parameter state unit is fused and modeled, including: Construct a candidate fusion state set on continuous time slices and establish a mapping relationship between the candidate fusion state set and the parameter state unit; A state consistency constraint model is introduced to perform constraint propagation on the candidate fusion state set, generating a set of feasible fusion paths that satisfy the constraint rule entries; Perform a consistency determination on the set of feasible fusion paths, select the fusion path with the fewest number of violation rule entries and the fewest state transition conflicts, and output the fusion state result corresponding to the fusion path.

8. The sleep multi-parameter fusion modeling method according to claim 7, characterized in that: The fusion state results are corrected based on anomaly detection mechanisms, including: A fusion state prediction window is established based on the fusion path. The prediction window is used to generate the predicted fusion state within the time slice. Calculate the consistency residual between the predicted fusion state and the fusion state result, and generate anomaly markers based on the consistency residual; When the anomaly marker meets the triggering condition, backtrack the time slice window corresponding to the anomaly marker and re-execute the candidate fusion state construction and constraint propagation, output the corrected fusion state result and record the correction window identifier.

9. The sleep multi-parameter fusion modeling method according to claim 8, characterized in that: The state consistency constraint model is updated, including: An updated sample entry is generated based on the corrected fusion state result. The updated sample entry includes time slice identifier, parameter state unit, fusion state result and anomaly marker. The number of times the constraint rule entries are satisfied and the number of times they are violated are statistically analyzed based on the updated sample entries, and the trigger condition boundaries of the constraint rule entries are adaptively adjusted. Incremental training is performed on the deep learning sequence modeling network based on updated sample entries, so that the relation representation vector adapts to the changes in the parameter state association relationship.

Images