Semantic adaptation method and device for multi-sampling rate data, edge device and storage medium

Abstract

The application discloses a kind of semantic adaptation methods of multiple sampling rate data, belong to data acquisition and edge computing technical field, to solve two layer sampling rate dimension misplacement, semantic loss and data expiration problem.Physiological characteristic data stream of perception end is received with first sampling rate and is written into annular buffer;According to feature type, automatically selected recent value keeps, time window mean, peak value keeps or quality weighted strategy;According to this, it is converted into with the decision feature of lower second sampling rate that aligns with the granularity of period;Additional effective period identifier containing sample time span start and end, sample number and effective cutoff time;With second sampling rate, the feature is output to decision layer, when it is beyond effective range, it is informed to switch to degraded decision logic;Two layer sampling rate is decoupled each other.The device includes feature stream access domain, strategy table, converger set, effective period manager and decision feature output domain module.Thereby eliminate interface suspension, reduce semantic loss and make data availability can be judged.

Description

Technical Field

[0001] This invention relates to the field of data acquisition and edge computing technology, and particularly to a semantic adaptation method, apparatus, edge device, and storage medium for multi-sampling rate data. This invention is applicable to data systems with two levels—high-frequency sampling and low-frequency decision-making—and where the two levels have different sampling rates, including but not limited to wearable and smart wearable devices, low-power IoT edge computing, millimeter-wave radar home monitoring, multi-rate industrial control systems, vehicle-to-everything (V2X) multi-source fusion, medical monitoring equipment, and smart home sensing hubs. Background Technology

[0002] In wearable, IoT, and physiological monitoring systems, the transition from high-frequency sampling data to low-frequency decision-making typically falls into one of six categories, each with its own drawbacks. The first category is direct downsampling. This approach extracts points from the high-frequency sequence at a fixed sampling ratio, or uses the last point within a window as the representative value for that window, achieving sampling rate transformation through extraction, resulting in minimal computation. Its limitations include: discarding all information within the window except for the sampling point; for features like body motion that are only meaningful at peak values, it might sample the resting point and miss a turning over; for features like heart rate that require averaging, single-point values ​​are easily affected by noise; it doesn't carry validity information, leaving downstream systems unable to determine the reliability of the value; and the sampling ratio is tied to the sampling rates of the two layers. The second category is unified window statistics. This approach performs the same statistical method on all samples within the window, typically an arithmetic mean, to suppress noise through averaging. Its limitations are: uniformity is its fundamental flaw; averaging body motion will smooth out a significant fluctuation, meaning multiple resting samples dilute a single peak; averaging heart rate variability will mask its temporal structure; stage labels, as discrete categories, cannot be averaged; arithmetic mean is sensitive to outliers and does not consider signal quality differences; and it also does not carry validity information. The third type is event-driven sampling. This scheme only records and reports when specific events occur, producing no data at other times, using sparse events instead of continuous sampling to save power and bandwidth. Its limitations are: no events does not equal no change; trend information during the silent period is completely lost; uneven event distribution leads to uncertain output time, making it impossible to align with downstream decision-making cycles; it cannot answer what the value of the feature is during the silent period; and it also does not carry validity information, as the silent period and the disconnected period appear identical to the downstream. The fourth type is a general-purpose streaming processing framework. This scheme uses a general-purpose streaming processing engine to process multi-rate data streams, providing windowing, watermarking, and out-of-order processing capabilities. Its limitations are: the water level line addresses out-of-order and late arrivals, not data expiration caused by upstream interruptions or underload; the window aggregation strategy is configured by operators rather than automatically selected based on feature semantics; the framework size and runtime overhead far exceed the capacity of the wearable device's microcontroller unit, conflicting with the goal of standby current not exceeding 50 microamps; and it does not provide a set of strategies oriented towards physiological feature semantics. The fifth category is unadapted synchronous blocking. This scheme allows the decision layer to be driven directly by high-frequency data or wait for a fixed window to fill before making a decision, with both layers sharing the same time base and buffer. Its limitations are: the sampling rates of the two layers are essentially bound, and any adjustment on one side will cause a chain reaction of modifications on the other side; if the decision layer is driven by high frequency, the computational load of phased interpretation and state transition increases by hundreds of times; if waiting for the window to fill, it will be permanently blocked when the sensor node is dormant or disconnected, with no timeout or degradation mechanism. This is precisely the cause of the misalignment between the 100 milliseconds and 30 seconds. The sixth category is general multi-rate signal processing. This scheme uses methods such as decimation, interpolation, anti-aliasing filtering, and multiphase filtering to achieve sampling rate transformation, and uses filtering to ensure that the sampling rate transformation is free from aliasing distortion.Its limitations are: it only solves the sampling rate transformation at the waveform level, and is not applicable to categorical features such as discrete period labels; it does not involve selecting different convergence methods according to semantics; it does not output time validity indicators; its goal is to reconstruct waveforms without distortion, while the goal of this scenario is to retain the information needed for decision-making.

[0003] Summarizing the above six types of solutions, the existing technologies have seven unresolved technical problems: misalignment of the two sampling rate dimensions leads to interface vacancy; unified downsampling causes semantic loss; downstream cannot determine whether data is expired; expired data and silent data are indistinguishable; binding of the two sampling rates makes independent evolution difficult; resource-constrained platforms have no available solutions; and there is a lack of clear basis for triggering downgrades.

[0004] This discrepancy stems from a misalignment in the system's internal metrics: the sensor architecture uses preprocessing latency of no more than 100 milliseconds, acquisition timing error of no more than 1 millisecond, and feature extraction and phased inference of no more than 50 milliseconds as immutable cross-layer metrics, while the decision layer uses 30-second epochs for four-state phased analysis and state transitions. Both layers' metrics are achievable, but when 100 milliseconds is used as the immutable cross-layer metric, the hardware team designs the buffer queue based on 100-milliseconds, while the software team subscribes to the feature stream based on 30-second intervals, resulting in an undefined interface.

[0005] In terms of the closest prior art, CN117158891A (Apple Inc., System and Method for Sleep State Tracking, Publication Date 2023-12-05) discloses a system and method for sleep state tracking. Its wearable device includes a motion tracking sensor and uses motion tracking sensor data to estimate and classify sleep states for multiple time periods to determine sleep intervals. It can also perform signal quality checks on motion tracking sensor data, preempt the display of sleep state classification or tracking results by the signal quality check, and filter and smooth the classified sleep states for each time period (see the abstract of the document, source: lt;). However, this document does not disclose the interface layer decoupling mechanism of "the first sampling rate feature stream and the second sampling rate decision feature are explicitly modeled as two independent quantities, and the sampling rate and time semantic conversion is completed by the adapter between the two layers", nor does it disclose the basis for "explicitly outputting the time validity information of the feature along with the decision feature and driving degradation according to the validity period identifier". CN115886770A (Apple Inc., Publication Date 2023-04-04) discloses a radar-based biometric signal extraction system and method. Its processing circuit performs different statistical analyses on the reflected signals from the monitored object and performs different processing functions based on the statistical analysis results to determine vibration, thereby extracting biometric information (such as respiratory and cardiovascular information) associated with the living object (see the abstract of this document, source: lt;). However, this document does not disclose the automatic selection of a convergence strategy among multiple convergence strategies (peak hold, time window mean, quality weighting, recent value hold or mode, count) according to the physiological characteristic type, nor does it involve semantic convergence of discrete categorical stage labels. CN106991810A (Hangzhou Dianzi University, Publication Date 2017-07-28) discloses a multi-sampling rate method applied to monitoring terminals. It achieves multi-sampling rate monitoring of vehicle exhaust by automatically adjusting the sampling rate of the monitoring terminal sensor, so as to eliminate data redundancy and improve the measurement accuracy of the equipment (see the abstract of the document, source: lt;). However, the document does not disclose explicitly modeling the two sampling rates as two independent quantities and setting an adapter between them, does not disclose selecting the convergence strategy according to the feature semantics, and does not disclose outputting the time validity identifier as the basis for judging expired data and triggering degradation.

[0006] The seven issues mentioned above are intertwined, all pointing to the lack of an explicit adaptation structure between the high-frequency sampling layer and the low-frequency decision layer. None of the aforementioned literature discloses a combination of "decoupling and modeling the first sampling rate feature stream and the second sampling rate decision feature, automatically selecting a convergence strategy based on physiological feature semantics, and explicitly carrying a time validity identifier along with the decision feature to drive degradation." Therefore, existing solutions generally lack the ability to decouple the interface layer when the two sampling rate dimensions are misaligned, and also lack a clear basis for distinguishing between expired data and silent data and triggering degradation accordingly. Summary of the Invention

[0007] The purpose of this invention is to provide a semantic adaptation method and apparatus for multi-sampling rate data, so as to solve the problems in the prior art such as misalignment of the two-layer sampling rate units, semantic loss caused by unified downsampling, inability of downstream to determine whether the data is expired, and difficulty in independent evolution caused by binding the two-layer sampling rates.

[0008] To achieve the above objectives, this invention provides a semantic adaptation method for multi-sampling-rate data, comprising the following steps: receiving at least one physiological feature data stream from a sensing end at a first sampling rate, wherein each feature frame contains a feature type identifier, feature value, signal quality index, and timestamp, and writing the feature frame into a circular buffer; selecting a corresponding convergence strategy for each of the at least one physiological feature data, wherein the convergence strategy includes at least one of recent value preservation, time window mean, peak preservation, and quality weighting, and the convergence strategy is automatically selected according to the type of physiological feature; and converting the physiological feature data at the first sampling rate into decision features at a second sampling rate, wherein the second sampling rate is lower than the first sampling rate. The first sampling rate and the second sampling rate are aligned with the decision-making granularity of the decision layer; an expiration date identifier is attached to the decision feature, which includes at least the start and end points of the time span of the original samples covered by the decision feature, the actual number of samples included in the aggregation, and the valid cutoff time; the decision feature carrying the expiration date identifier is output to the decision layer at the second sampling rate; when the decision feature exceeds the valid range represented by the expiration date identifier, the decision layer is notified to switch to the downgrade judgment logic, instead of continuing to judge with expired data or directly judging failure; the first sampling rate and the second sampling rate are decoupled from each other, changing the first sampling rate does not affect the second sampling rate, and changing the second sampling rate does not affect the first sampling rate. The above technical solution is located in the four-layer architecture of the ubiquitous IoT sleep management system, between the sensor perception layer and the data management and analysis layer, that is, between the feature flow layer and the decision flow layer. It does not perform sensing, feature extraction, sleep stage interpretation and rhythm modeling, but only completes the conversion between sampling rate and temporal semantics. Vertically, this method receives the first sampling rate feature flow output by heterogeneous sensor nodes after scenario scheduling, and provides decision features for the hierarchical processing state machine and hard cutoff guards. In terms of product form, this method can be an adaptation module running on the firmware of the main control microcontroller unit of wearable devices, middleware running on the bedside central control or home gateway, or a two-segment adaptation distributed between sensor nodes and the host.

[0009] Furthermore, vertically, this method receives the first sampling rate feature stream output by heterogeneous sensor nodes after scenario-based scheduling, and provides decision features for the hierarchical processing state machine and hard cutoff guards. Horizontally, it directly connects with the degradation inference chain of disconnection scenarios; the validity period identifier output by this method is the basis for determining whether data has expired and whether degradation is necessary. In terms of product form, this method can exist in three forms: an adaptation module running on the firmware of the wearable device's main control microcontroller unit, middleware running on a bedside central control unit or home gateway, and a two-segment adaptation distributed between sensor nodes and the host. This method only involves the interface layer mechanism of dual sampling rate decoupling, semantic convergence strategy, and validity period-driven degradation; it does not involve phased algorithms, rhythm modeling, or general downsampling and interpolation techniques.

[0010] Furthermore, the above steps constitute a complete causal chain: decoupling provides the structure, the semantic strategy ensures that downsampling is not distorted, the expiration date provides temporal semantics, degradation ensures robustness, and configurability and lightweight implementation ensure engineering feasibility. Among them, the semantic convergence strategy solves the problem of which part of the information to retain during downsampling, and the expiration date identifier solves the problem of when the output value is no longer reliable. The two work together to reduce semantic loss and trigger degradation.

[0011] Furthermore, further improvements of the present invention include: explicitly modeling the sampling rates of the two layers as decoupled quantities to eliminate misalignment of dimensional hierarchy; automatically selecting convergence strategies based on the semantics of physiological characteristics to avoid semantic loss caused by uniform downsampling; attaching validity period identifiers to decision features so that downstream components can determine whether the data has expired; using validity period determination to drive the decision layer to switch to degradation logic instead of using expired data; implementing the scheme with circular buffering and incremental accumulation so that the scheme can run on resource-constrained edge microcontroller units; and making the strategy table configurable so that adding or deleting feature types does not require code modification.

[0012] Furthermore, the simplified implementation means that at edge devices with limited computing power and storage, the two items in the validity period identifier that need to be obtained by local real-time accumulation—the actual number of samples and the valid cutoff time—are directly written by the edge device, while the three items that need to rely on global information—the start and end points of the time span, the expected number of samples, and the coverage rate—are supplemented and written by the host device when the configuration is issued. When this implementation is adopted, the edge device only maintains a circular buffer and two counters, and does not save the complete time series.

[0013] Beneficial effects

[0014] Compared with existing technologies, the present invention has the following advantages: First, the two sampling rates are explicitly decoupled, the interface dangling is eliminated, and the two layers can evolve independently; Second, the semantic selection strategy ensures that the peak value of body movement, the mean value of heart rate, the quality weighting of heart rate variability, and the mode of the stage label are all properly assigned, reducing information loss; Third, the validity period identifier makes the usability of data a determinable quantity, and disconnection and normal dormancy can be distinguished; Fourth, degradation is explicitly triggered by the validity period determination point, and the timing and extent of degradation are based on evidence; Fifth, the circular buffer plus incremental accumulation can run on the wearable microcontroller unit; Sixth, the strategy table is configurable, and the addition or deletion of feature types does not require code modification. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Figure 2 This is a diagram of the overall structure of the dual-sampling-rate feature-decision adapter (also serving as a device structure block diagram). Figure 3 A time series diagram of dual sampling rate decoupling and incremental convergence within the decision cycle; Figure 4 A comparison chart of semantically selected convergence strategies for various features and uniform downsampling; Figure 5 This diagram illustrates the determination of expiration date identification and downgrade triggering. The above figures are for illustrative purposes only, illustrating the structure and process of the present invention, and are not intended to limit the scope of protection of the present invention.

[0016] Figure 1 Explanation: This shows the execution order of steps S1 to S7, as well as the two branches: the decision node for whether the decision feature is still valid and the downgrade decision logic.

[0017] Figure 2 Explanation: The diagram shows the top-to-bottom sequence of the feature stream access domain, ring buffer, policy table module, aggregator collection module, validity manager module, and decision feature output domain module.

[0018] Figure 3 Explanation: This shows the timing of writing the first sampling rate into the ring buffer, the aggregate maintaining the incremental state, and the second sampling rate outputting decision features and validity period identifiers.

[0019] Figure 4 Note: By comparing the semantically self-selected convergence strategies for various features with the consequences of uniform downsampling, the semantic loss caused by uniform downsampling is shown for each category.

[0020] Figure 5 Note: This shows the determination of three scenarios: under-coverage, later than the effective deadline, and no samples arriving for several consecutive decision periods. If all three scenarios are negative, the decision feature is valid. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. The following uses a wearable sleep monitoring process as an example to illustrate the workflow and effects of this invention; the given values ​​are exemplary and do not constitute a limitation on the scope of protection of this invention. The flow of the method of this invention is as follows: Figure 1 As shown, the specific implementation methods for each step are as follows.

[0022] Step S1: Receive at least one physiological feature data stream from the sensing end at a first sampling rate. Each feature frame contains a feature type identifier, feature value, signal quality index and timestamp, and write the feature frame into a circular buffer.

[0023] In step S1, the capacity of the ring buffer is determined by the first sampling rate and the maximum tolerable delay, and is independent of the second sampling rate. All data structures are statically allocated during initialization and no dynamic memory allocation is performed during runtime. In this embodiment, the feature stream access domain receives feature frames from the sensing end at the first sampling rate f1. Each feature frame contains a feature type identifier, feature value, signal quality index, and timestamp. The access domain does not perform any downsampling, only buffering and timestamp alignment. The physiological feature data stream accessed by this method includes at least one of photoplethysmography (PPG), triaxial acceleration, heart rate, heart rate variability, blood oxygen saturation, respiratory rate, sound intensity, ambient light, and sleep stage labels. PPG is the source of blood oxygen saturation and pulse wave morphology; sound intensity reflects environmental and respiratory acoustic characteristics; and ambient light reflects the ambient light environment. In this embodiment, the ring buffer depth is designed based on the first sampling rate and the maximum tolerable delay, and is independent of the second sampling rate. The convergent state capacity is designed based on the feature type, and is independent of the first sampling rate. The position of the feature stream access domain and the ring buffer in the overall structure of this adapter and their data stream relationship are as follows: Figure 2 As shown.

[0024] Step S2: Select a corresponding convergence strategy for each of the at least one physiological feature data. The convergence strategy includes at least one of the following: recent value preservation, time window mean, peak preservation, and quality weighting. The convergence strategy is automatically selected according to the type of physiological feature.

[0025] In step S2, the convergence strategy is automatically selected based on the type of physiological characteristic. Specifically, this includes: peak hold strategy for body movement characteristics, time window mean strategy for heart rate and respiratory rate characteristics, quality weighting strategy for heart rate variability and blood oxygenation characteristics, recent value hold or mode strategy for stage label and posture category characteristics, and counting strategy for event-type characteristics. The convergence strategy is configured in the form of a strategy table, which records the mapping from feature type identifier to convergence strategy identifier and its parameters. When a new feature type is added, only one record is added to the strategy table without modifying the adapter's program code. In this embodiment, the strategy table records the convergence strategy and its parameters corresponding to each feature type, is configured in data form, and can be updated over the air or locally. When the adapter accesses a feature stream, it looks up the strategy in the table according to its type identifier, without needing to write branch code for each feature. The key difference between this invention and uniform downsampling lies in the automatic selection of convergence strategies based on semantic features: the key to body movement features is whether the event has occurred, so peak preservation is chosen; if time window mean is used instead, a significant turning over will be diluted and smoothed out by a large number of resting samples; the key to heart rate and respiratory rate is the average level, so time window mean is chosen; if peak preservation is used instead, it will be affected by transient noise; heart rate variability and blood oxygen are highly sensitive to signal quality, so quality weighting is chosen to deweight low-quality samples; stage labels and posture categories are discrete categorical features with no average meaning, so recent value preservation or mode is chosen. The calculation methods for each convergence strategy are as follows: Let M be the number of samples within the window. The most recent value is the latest sample value. The time window mean is the average of the samples within the window; a truncated mean can be selected, removing extreme values ​​at both ends. The peak value is the maximum value of the sample values ​​within the window, or the maximum absolute value of the sample values. Quality weighting uses signal quality as the weight and calculates the weighted average of the sample values; the signal quality index of the samples has been normalized to the zero-to-one range. The mode is the category that appears most frequently within the window. Counting is used to count the number of samples exceeding the threshold within the window; the threshold is the threshold for event-type features, and the frequency of the category within the window is used for the mode. The above convergence strategies are not uniform; different types of first and second features exist, leading to different convergence strategies for each. A comparison of the four types of features with semantically self-selected convergence strategies and uniform downsampling is shown below. Figure 4 As shown.

[0026] Step S3: Based on the convergence strategy, the physiological feature data of the first sampling rate is converted into decision features of the second sampling rate. The second sampling rate is lower than the first sampling rate, and the decision cycle corresponding to the second sampling rate is aligned with the stage granularity of the decision layer.

[0027] In step S3, the length of the decision period is automatically derived from the granularity parameter, which gives the length of the epoch. The decision period takes this length, and the second sampling rate is the reciprocal of this length. The time bases for the first and second sampling rates are maintained independently, and a mapping is performed between the millisecond timestamp of the physiological feature data stream and the epoch number or absolute time of the decision layer. Each aggregator only maintains the incremental state of constant order. The incremental state includes at least one of the following: cumulative sum and count, current maximum value, weighted numerator and weighted denominator, and category count table. When a new sample arrives, only the incremental state is updated; the original samples of the entire window are not saved, making memory usage independent of the window length. In this embodiment, the first sampling rate is f1, in Hertz, corresponding to the feature frame rate; the second sampling rate is f2, in Hertz, corresponding to the decision frame rate; and the decision period is equal to the reciprocal of the second sampling rate. In this implementation, dual sampling rate decoupling has four meanings: value decoupling, meaning that the first sampling rate and the second sampling rate are independently valued, and their ratio is not a compile-time constant but a runtime variable; time base decoupling, meaning that the feature flow layer uses millisecond timestamps as the base, and the decision flow layer uses the epoch number or absolute time as the base, with the adapter performing mapping between the two; derivable decision cycle, meaning that the decision cycle length is automatically derived from the epoch granularity parameter, and the adapter sets the second sampling rate accordingly without manual secondary configuration, resolving the problem of the decision cycle being confused with preprocessing latency; and buffer decoupling, meaning that the ring buffer depth is designed according to the first sampling rate and the maximum tolerable latency and is independent of the second sampling rate, and the aggregator state capacity is designed according to the feature type and is independent of the first sampling rate. In this embodiment, the first sampling rate f1 is 50 Hz, corresponding to the hard real-time link of the perception layer, with a preprocessing delay of no more than 100 milliseconds; the epoch length is 30 seconds, so the decision period is 30 seconds; the second sampling rate f2 is one-thirtieth of a Hz; during settlement, the body motion convergent uses the maximum value of the acceleration magnitude within the window, the heart rate convergent uses the truncated mean of 1500 samples, the heart rate variability convergent uses a weighted average based on signal quality, and the epoch label convergent uses the category with the most occurrences. The temporal relationship between dual sampling rate decoupling and incremental convergence within the decision period is as follows: Figure 3 As shown.

[0028] Step S4: Add a validity period identifier to the decision feature. The validity period identifier shall at least include the start and end points of the time span of the original samples covered by the decision feature, the actual number of samples included in the aggregation, and the valid cutoff time.

[0029] In step S4, the validity period identifier also includes the expected number of samples and the coverage rate. The expected number of samples is determined by the product of the first sampling rate and the decision period. The coverage rate is determined by the ratio of the actual number of samples to the expected number of samples. When the coverage rate is lower than a preset threshold, it is determined to be underloaded and an underload flag is set. The validity period of the validity period identifier is a positive integer multiple of the decision period and is not greater than the tolerance period of physiological characteristics in downstream decision-making. In this embodiment, the validity period manager writes the validity period identifier at the end of each decision period and determines whether it is still within the validity period when querying downstream. The start and end points of the time span in the validity period identifier are the timestamps of the first and last samples of the period, respectively. The actual number of samples is the number of samples actually included in the aggregation. The expected number of samples is the product of the first sampling rate and the decision period. The coverage rate is the ratio of the actual number of samples to the expected number of samples. The valid expiration time is equal to the end point of the time span plus the valid duration. The effective duration being a positive integer multiple of the decision cycle means that the effective duration equals the product of a positive integer and the decision cycle, where the positive integer is not less than 1. This method aligns the decision cycle with the validity period boundary, preventing the validity period from expiring midway through the decision cycle. In this embodiment, the effective duration is twice the decision cycle, i.e., 60 seconds, and the coverage threshold is 0.8; the expected sample size is 50 × 30, i.e., 1500, and no underload flag is set when the coverage is not less than 0.8. The positions of the validity period manager and the decision feature output domain in the overall adapter structure are as follows: Figure 2 As shown.

[0030] Step S5: Output the decision features carrying the validity period identifier to the decision layer at the second sampling rate;

[0031] In step S5, the decision feature is output to the hierarchical processing state machine as the criterion for state transition. The coverage rate and underload flag in the validity period identifier are output to the degradation inference chain as the basis for degradation triggering. This method does not participate in state transition decision-making, nor does it execute the degradation inference algorithm. In this embodiment, the decision feature output domain outputs the decision feature carrying the validity period identifier at the second sampling rate f2, and outputs a degradation notification when the validity period expires. The output decision feature carries the value of the original sample after aggregation within the decision period, and its time base is the epoch number or absolute time of the decision layer. The alignment relationship between the output of the decision feature and the decision period for the second sampling rate is as follows: Figure 3 As shown.

[0032] Step S6: When the decision feature exceeds the valid range represented by the validity period identifier, notify the decision layer to switch to the downgrade judgment logic, instead of continuing to judge with expired data or directly judging failure.

[0033] In step S6, exceeding the valid range includes the following three scenarios: Underload scenario: the coverage rate is lower than a preset threshold, and the underload flag indicates that the underload originated from normal sleep. The decision layer uses this decision feature with low confidence and is not allowed to trigger an improved state transition based on it. Expiration scenario: the current time is later than the valid expiration time, and the decision layer is notified to switch to the degradation judgment logic. Interruption scenario: no samples arrive for several consecutive decision periods, and underload is directly marked and a degradation notification is output without waiting for the valid expiration time to arrive. In this embodiment, when the access domain detects that no new feature frames arrive consecutively or the number of samples in the current period is lower than the lower limit, the validity period manager directly marks the current period as underload and outputs a degradation notification. This bypass is the core path in the disconnection scenario. After receiving the decision feature, the decision layer first compares the current time with the valid deadline: if it is not overdue and the coverage rate meets the standard, it is used normally and classified into four states accordingly; if the coverage rate is lower than 0.8 but the underload flag indicates that the underload is due to normal dormancy, the decision feature is used with low confidence, maintaining the current state but not allowing the triggering of an improved migration; if the current time is later than the valid deadline, the degradation judgment logic is switched, and the decision feature is no longer used for judgment. As a contrast, when a sensor node disconnects, the coverage rate for that period is significantly lower than 0.8 and is marked as underloaded. The decision layer switches to degradation logic accordingly and will not continue to use old data for judgment after disconnection. The judgment node in this step and the corresponding handling paths for the two judgment results are as follows: Figure 1 As shown. The three situations exceeding the effective range in this step and their handling are as follows: Figure 5 As shown.

[0034] In step S7, the first sampling rate and the second sampling rate are decoupled from each other. Changing the first sampling rate does not affect the second sampling rate, and changing the second sampling rate does not affect the first sampling rate.

[0035] In step S7, this method is implemented in one of the following three forms: an adaptation module running on the firmware of the wearable device's main control microcontroller unit; middleware running on the bedside central control or home gateway; or a two-segment adaptation distributed between the sensor node and the host. In this embodiment, the first sampling rate f1 and the second sampling rate f2 are independent variables, and their ratio is a runtime variable. Changing the first sampling rate does not change the second sampling rate, and changing the second sampling rate does not change the first sampling rate. When the first sampling rate is increased from 50 Hz to 100 Hz to improve motion detection, the second sampling rate remains unchanged. When the decision cycle is changed from 30 seconds to 20 seconds, and the second sampling rate is correspondingly changed from 1 / 30 Hz to 1 / 20 Hz to refine the granularity of the epochs, the first sampling rate remains unchanged. Its engineering value lies in the fact that the hardware team only needs to be responsible for the first sampling rate and the metric of no more than 100 milliseconds, and the software team only needs to be responsible for the second sampling rate and the semantics of the epochs. The interface contract is explicitly carried by the adapter. The decoupling relationship between the first sampling rate and the second sampling rate and the incremental convergence timing within the decision cycle are as follows: Figure 3 As shown.

[0036] This embodiment also provides a semantic adaptation device for multi-sampling rate data, including a feature stream access domain, a policy table module, an aggregator set module, a validity period manager module, and a decision feature output domain module: the feature stream access domain is used to receive feature frames at a first sampling rate and write them to a circular buffer; the policy table module is used to record the mapping from feature types to aggregation policies in data form; the aggregator set module is used to maintain the incremental state for each feature according to the policy and settle it in the decision cycle; the validity period manager module is used to write the validity period identifier and determine whether it has expired or is underloaded; the decision feature output domain module is used to output the decision feature carrying the validity period identifier at a second sampling rate and output a degradation notification when it expires. In this embodiment, the above modules are implemented by firmware in the same wearable device main control microcontroller unit, and all data structures are statically allocated during initialization and there is no heap allocation during runtime. The structure of the device is as follows: Figure 2 As shown.

[0037] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0038] The above embodiments are only used to illustrate the design concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made based on the principles and design ideas disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A semantic adaptation method for multi-sampling rate data, characterized in that, The process includes the following steps: Step S1, receiving at least one physiological feature data stream from the sensing end at a first sampling rate, each feature frame containing a feature type identifier, feature value, signal quality index, and timestamp, and writing the feature frame into a circular buffer; Step S2, selecting a corresponding convergence strategy for each of the at least one physiological feature data, the convergence strategy including at least one of recent value preservation, time window mean, peak preservation, and quality weighting, and the convergence strategy is automatically selected according to the type of physiological feature; Step S3, converting the physiological feature data at the first sampling rate into decision features at a second sampling rate according to the convergence strategy, the second sampling rate being lower than the first sampling rate, and the decision period corresponding to the second sampling rate being... The granularity of the period is aligned with that of the decision-making layer; Step S4: Attach an expiration date identifier to the decision feature. The expiration date identifier includes at least the start and end points of the time span of the original samples covered by the decision feature, the actual number of samples included in the aggregation, and the valid cutoff time; Step S5: Output the decision feature carrying the expiration date identifier to the decision-making layer at the second sampling rate; Step S6: When the decision feature exceeds the valid range represented by the expiration date identifier, notify the decision-making layer to switch to the downgrade judgment logic, instead of continuing to judge with expired data or directly judging failure; Step S7: The first sampling rate and the second sampling rate are decoupled from each other. Changing the first sampling rate does not affect the second sampling rate, and changing the second sampling rate does not affect the first sampling rate.

2. The semantic adaptation method for multi-sampling rate data according to claim 1, characterized in that, In step S2, the convergence strategy is automatically selected based on the type of physiological feature. Specifically, the following strategies are used: peak retention strategy for body movement features, time window mean strategy for heart rate and respiratory rate features, quality weighting strategy for heart rate variability and blood oxygenation features, recent value retention or mode strategy for stage label and posture category features, and counting strategy for event-type features. In step S2, the convergence strategy is configured in the form of a strategy table. The strategy table records the mapping from feature type identifier to convergence strategy identifier and its parameters. When a new feature type is added, only one record is added to the strategy table without modifying the adapter program code.

3. The semantic adaptation method for multi-sampling rate data according to claim 2, characterized in that, In step S3, the length of the decision cycle is automatically derived by the granularity parameter of the periodization, which gives the length of the periodization epoch. The decision cycle takes this length, and the second sampling rate takes the reciprocal of this length. The time bases of the first sampling rate and the second sampling rate are maintained independently, and a mapping is performed between the millisecond timestamp of the physiological feature data stream and the epoch number or absolute time of the decision layer.

4. The semantic adaptation method for multi-sampling rate data according to claim 1, characterized in that, In step S4, the validity period identifier also includes the expected number of samples and the coverage rate. The expected number of samples is determined by the product of the first sampling rate and the decision period. The coverage rate is determined by the ratio of the actual number of samples to the expected number of samples. When the coverage rate is lower than a preset threshold, it is determined to be underloaded and an underload flag is set.

5. The semantic adaptation method for multi-sampling rate data according to claim 4, characterized in that, Step S6 includes the following three scenarios where the effective range is exceeded: Underload scenario, where the coverage rate is lower than the preset threshold and the underload flag indicates that the underload is due to normal dormancy, the decision layer uses the decision feature with low confidence and is not allowed to trigger an improved state transition based on it; Expiration scenario, where the current time is later than the effective deadline, the decision layer is notified to switch to the degradation judgment logic; Interruption scenario, where no samples arrive within several consecutive decision cycles, underload is directly marked and a degradation notification is output without waiting for the effective deadline to arrive.

6. The semantic adaptation method for multi-sampling rate data according to claim 4, characterized in that, In step S3, each aggregator maintains only the incremental state of constant order. The incremental state includes at least one of the following: cumulative sum and count, current maximum value, weighted numerator and weighted denominator, and category count table. When a new sample arrives, only the incremental state is updated, and the original samples of the entire window are not saved, so that memory usage is independent of the window length. In step S1, the capacity of the ring buffer is determined according to the first sampling rate and the maximum tolerable delay and is independent of the second sampling rate. All data structures are statically allocated during initialization and no dynamic memory allocation is performed during runtime.

7. The semantic adaptation method for multi-sampling rate data according to claim 1, characterized in that, The validity period of the validity period identifier shall be a positive integer multiple of the decision-making cycle, and shall not exceed the tolerance period of physiological characteristics in downstream decision-making.

8. The semantic adaptation method for multi-sampling rate data according to claim 1, characterized in that, The decision features are output to the hierarchical processing state machine as the source of state transition criteria. The coverage rate and underload flag in the validity period identifier are output to the degradation inference chain as the basis for degradation triggering. This method does not participate in state transition decision-making, nor does it execute the degradation inference algorithm. This method is implemented in one of the following three forms: an adaptation module running on the firmware of the wearable device's main control microcontroller unit, middleware running on the bedside central control or home gateway, or a two-segment adaptation distributed between the sensor node and the host.

9. A semantic adaptation device for multi-sampling rate data, characterized in that, include: The feature stream access domain is used to receive feature frames at a first sampling rate and write them to a circular buffer. The strategy table module records the mapping from feature types to aggregation strategies in data form; the aggregator set module maintains the incremental status for each feature according to the strategy and settles it in the decision cycle; the validity period manager module writes the validity period identifier and determines whether it has expired or is underloaded. The decision feature output domain module is used to output decision features carrying validity period identifiers at a second sampling rate and output a downgrade notification when the validity period expires.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 8.

Images