Power supply distribution method of wearable device, electronic device, and storage medium

Abstract

The present application relates to the technical field of wearable equipment, in particular to a power supply distribution method of a wearable device, an electronic device and a storage medium, the method comprises: adjusting real-time position coordinates of a plurality of detectors in the wearable device, and comparing with standard position coordinates to obtain offset information of each detector; inputting current monitoring data of the detector and the offset information into a prediction model to output prediction data with a confidence label, and calibrating the prediction data through an adjustment model to obtain adjusted monitoring data; based on the adjusted monitoring data and current information of a user, identifying abnormal monitoring data and a corresponding abnormal detector; when it is monitored that the current remaining power of the wearable device is low, adjusting the power supply and collection frequency of non-abnormal detectors until the current remaining power recovers to be higher than a second preset power threshold; the present application can improve the rationality of power supply distribution, realize dynamic balance of core function guarantee and precise resource allocation for prolonging the endurance.

Description

Technical Field

[0001] This invention relates to the field of wearable device technology, and in particular to a power distribution method, electronic device, and storage medium for wearable devices. Background Technology

[0002] With the widespread adoption of wearable health monitoring devices, improving monitoring accuracy and achieving intelligent power management without affecting core monitoring functions have become key challenges for enhancing user experience and device reliability. Current wearable devices typically report data in real-time. However, due to the tendency for wearable devices to shift positions during use, the accuracy of sensor data decreases, directly impacting the accuracy of subsequent health status assessments. Furthermore, current wearable devices often employ a unified power supply strategy based on fixed thresholds or simple battery percentages, synchronously reducing the frequency or shutting down all sensor modules when the battery is low. While this method can extend battery life, it fails to strike a balance between energy saving and ensuring core real-time functionality, potentially leading to the loss of critical monitoring data and missed detections of anomalies. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a power distribution method, electronic device, and storage medium for wearable devices, which can improve the accuracy of data monitoring and the rationality of power distribution, and achieve a dynamic balance between ensuring core functions and extending battery life through precise resource allocation.

[0004] According to a first aspect of the present invention, a power distribution method for a wearable device is provided, comprising the following steps:

[0005] S1 adjusts the real-time position coordinates of several detectors in the wearable device and compares them with the standard position coordinates of several pre-stored detectors to obtain offset information representing the offset state of each detector; where different detectors correspond to different body parts of the user.

[0006] S2, for any detector, input the current monitoring data and corresponding offset information of the detector into the pre-trained prediction model, and output the prediction data with confidence labels; the confidence labels are negatively correlated with the offset information.

[0007] S3 inputs the offset information, prediction data and corresponding confidence labels of each detector into the pre-trained adjustment model, calibrates the prediction data based on the confidence labels, and outputs the adjusted monitoring data.

[0008] S4 generates model prompts based on the user's current motion status information and preset user profile features, and identifies abnormal monitoring data and corresponding abnormal detectors based on the adjusted monitoring data and the preset abnormal detection model.

[0009] S5, when the current remaining power of the wearable device is detected to be lower than the first preset power threshold, the power supply and sampling frequency of each non-abnormal detector except the abnormal detector are adjusted until the current remaining power is restored to a level higher than the second preset power threshold; the second preset power threshold is greater than the first preset power threshold.

[0010] According to a second aspect of the present invention, a non-transitory computer-readable storage medium is provided, wherein at least one instruction or at least one program is stored therein, the at least one instruction or the at least one program being loaded and executed by a processor to implement the power distribution method for the wearable device described above.

[0011] According to a third aspect of the present invention, an electronic device is provided, including a processor and the aforementioned non-transitory computer-readable storage medium.

[0012] The present invention has at least the following beneficial effects:

[0013] This invention provides a power distribution method for wearable devices. First, the real-time position coordinates of several detectors in the wearable device are adjusted and compared with standard position coordinates to obtain the offset information of each detector. This provides factual basis for subsequent data monitoring and adjustment, improving the accuracy of data monitoring. Then, the current monitoring data and offset information of the detectors are input into a prediction model, which outputs predicted data with confidence labels. This predicted data is then calibrated using an adjustment model to obtain more accurate monitoring data. By introducing confidence labels and assigning different weight coefficients to different confidence labels, the adjustment model can prioritize the calibration of low-confidence monitoring data, providing highly reliable calibration for subsequent anomaly detection models. The accurate post-data significantly reduces the false judgment rate caused by data distortion; based on the adjusted monitoring data and the user's current information, abnormal monitoring data and corresponding abnormal detectors are identified. By integrating dynamic real-time behavior with static individual characteristics, the accuracy of abnormal data monitoring is improved; finally, when the wearable device's current remaining power is detected to be low, the power supply and collection frequency of non-abnormal detectors are adjusted until the current remaining power is restored to above the second preset power threshold. Differentiated frequency reduction and voltage adjustment strategies are implemented according to different detector priorities, achieving a dynamic balance between precise resource allocation and core function protection, significantly improving the wearable device's ability to maintain function and user experience in low power conditions. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 A flowchart illustrating a power distribution method for a wearable device provided in an embodiment of the present invention. Detailed Implementation

[0016] 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.

[0017] This invention provides a power distribution method for wearable devices, such as... Figure 1 As shown, the method includes the following steps:

[0018] S1 adjusts the real-time position coordinates of several detectors in the wearable device and compares them with the standard position coordinates of several pre-stored detectors to obtain offset information characterizing the offset state of each detector; this can be understood as: the real-time position coordinates of the detector are the real-time acquired spatial three-dimensional coordinates.

[0019] Specifically, different detectors correspond to different parts of the user's body. In practice, wearable devices can be flexible devices, and detectors can be detection units positioned for areas such as the chest, heart, and abdomen.

[0020] In a specific embodiment, step S1 includes the following steps:

[0021] S101, acquire device attitude data collected by inertial measurement units set in different local areas of the wearable device; this can be understood as: there is at least one detector in each local area.

[0022] S102, based on a predefined device region model, device posture data, and human kinematics model, establishes a dynamic transformation relationship from the device body coordinate system of each detector's local region to a preset standard human reference coordinate system. This can be understood as establishing a corresponding device body coordinate system for each local region. In specific implementation, under standard posture, through optical motion capture or manual measurement, the fixed transformation relationship between each local device body coordinate system and adjacent human skeletal segments is calibrated, achieving a basic connection between the device and the human body. Then, the entire device is treated as a physical mesh, with each detector position corresponding to a mesh node. Device posture data is used as node constraints, combined with the human kinematics model as the driving skeleton, to run a simplified physical simulation in real time, calculating the coordinates of each detector position in the standard human reference coordinate system. Furthermore, through extensive motion capture, a parameterized neural network model can be learned. This model takes human joint angles and device posture data collected by the inertial measurement unit as input, directly outputting the transformation matrix from the device body coordinate system to the standard human reference coordinate system for each local region.

[0023] S103, based on the dynamic transformation relationship, the actual measured coordinates of each detector in the corresponding device body coordinate system are transformed to the standard human body reference coordinate system to obtain the normalized actual position coordinates of each detector.

[0024] S104, In the standard human body reference coordinate system, the normalized actual position coordinates of each detector are compared with the pre-stored theoretical standard position coordinates of each detector to calculate the offset information of each detector; the offset information includes at least the offset distance. The offset distance is obtained by calculating the Euclidean distance of the spatial difference.

[0025] Preferably, the offset information also includes the offset type, such as vertical offset, horizontal offset, rotational offset, etc.

[0026] As mentioned above, the spatial position of the detector changes when the user is in motion, but the displacement relative to the original detection site may be very small. For example, when the user bends over, the position of the detector will drop but will still be in front of the chest. Therefore, it is necessary to map the real-time position coordinates of the detector to the standard position coordinates to obtain accurate offset information of the detector relative to the detection site. This provides factual basis for subsequent data monitoring and adjustment and helps to improve the accuracy of data monitoring.

[0027] S2, for any detector, input the current monitoring data and corresponding offset information of the detector into the pre-trained prediction model, and output the prediction data with confidence labels; the confidence labels are negatively correlated with the offset information; it can be understood that: the larger the offset distance, the smaller the confidence label of the prediction data, and the confidence label refers to the confidence score or confidence level.

[0028] Specifically, the training process of the prediction model is as follows:

[0029] S201, construct a first training sample set corresponding to several detectors; each training sample in the first training sample set includes an offset information sample and a corresponding original monitoring data sample, as well as a standard data sample collected through a preset reference device; it can be understood that the preset reference device can be a high-precision detection device that is pasted on the body part during the experiment. In implementation, the first training sample set is divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0030] S202, construct a target machine learning model with dual output branches and a confidence mapping function; the dual output branches are used to generate prediction data and intermediate parameters that characterize prediction uncertainty, respectively; for example, the intermediate parameters may be the prediction variance corresponding to the prediction data.

[0031] In one specific embodiment, the target machine learning model adopts an encoder-decoder structure. The encoder uses a one-dimensional convolutional neural network as a feature extractor. Its input layer receives the concatenated raw monitoring data samples and offset information samples. The first and second convolutional layers each have 3 kernels with a stride of 1, and the activation function is ReLU. The number of filters is 64 and 128, respectively. The pooling window size of the pooling layer is 2, with a stride of 2. The fully connected layer has 256 neurons, and the activation function is ReLU. The decoder contains two independent output branches. The prediction data branch uses a three-layer fully connected network with 128 and 64 neurons, and an output dimension consistent with the standard data sample dimension, using the ReLU activation function. The output layer has no activation function and is used to generate prediction data. The intermediate parameter branch uses a two-layer fully connected network with 32 and 1 neurons, respectively, and the activation function is ReLU. It outputs a scalar parameter greater than 0 to characterize the uncertainty of the prediction.

[0032] Specifically, the confidence mapping function is used to map the output intermediate parameters into confidence scores; this can be understood as mapping the uncertain estimated value of the model output to a confidence score between 0 and 1 through a monotonically decreasing function. For example, an exponential decay function is used to convert the intermediate parameters into confidence scores, and the calculation formula is c=exp(-α×σ), where α is a preset scaling factor, which is taken as 0.5 in this embodiment, and σ is the uncertain parameter value output by the intermediate parameter branch.

[0033] S203: Input each offset information sample and its corresponding original monitoring data sample into the target machine learning model. Use the standard data sample corresponding to each offset information sample as the supervision target. Update the parameters of the target machine learning model by optimizing the preset joint loss function to obtain the prediction model. In specific implementation, the model parameters are updated using gradient descent based on the calculated loss value. The gradient clipping threshold can be set to 1. Repeat the above steps until the validation set loss value converges or the maximum number of training rounds is reached.

[0034] Specifically, the joint loss function is constructed by weighting a data reconstruction loss term and an uncertainty regularization loss term. The data reconstruction loss term is used to calculate the difference between the predicted data and the standard data sample, while the uncertainty regularization loss term is used to encourage the model to learn to output a larger uncertainty estimate when the offset is greater and the prediction error is greater, that is, to output a lower confidence score.

[0035] Specifically, the formula for calculating the joint loss function L1 is as follows:

[0036] Where N is the number of the first training samples in this training, and M is the dimension of the prediction data. For the predicted data of the i-th first training sample, Let η be the standard data for the i-th first training sample. i λ is the intermediate parameter corresponding to the i-th first training sample, and λ is a preset weight coefficient used to balance the reconstruction accuracy and the stability of the uncertain estimation; in this embodiment, λ is set to 0.01.

[0037] As mentioned above, during the training of the prediction model, optimization was performed using a dual-output structure and a specific joint loss function, enabling the prediction model to simultaneously output corrected data and quantified confidence scores. By outputting predicted data and confidence scores, the prediction model not only corrected the original monitoring data, making it closer to the user's current real data, but also assigned confidence scores, providing a quantitative basis for subsequent data adjustments and power supply management responses, thereby improving the rationality and reliability of subsequent power supply allocation.

[0038] S3 inputs the offset information, prediction data and corresponding confidence labels of each detector into the pre-trained adjustment model, calibrates the prediction data based on the confidence labels, and outputs the adjusted monitoring data.

[0039] Specifically, the training process of the adjusted model is as follows:

[0040] S301, Construct a second training sample set; each training sample in the second training sample set includes an offset information sample, a corresponding standard data sample, and prediction data with confidence labels corresponding to each offset information sample output by the prediction model. In implementation, the second training sample set is divided into a training set, a validation set, and a test set in an 8:1:1 ratio. A stratified sampling strategy is used during the partitioning to ensure that the distribution ratio of different offset intervals and different confidence intervals is consistent across the datasets.

[0041] S302, the offset information samples and the predicted data with confidence labels corresponding to each offset information sample output by the prediction model are used as input features, and the corresponding standard data samples are used as supervision targets to iteratively optimize the weighted loss function set in the preset neural network model; wherein, the weighted loss function represents that the weight of each training sample is negatively correlated with the confidence label; it can be understood that: when calculating the calibration loss between the model output value and the sample true value, a higher loss weight coefficient is set for low confidence samples in the calibration loss function so that the model prioritizes optimizing the calibration accuracy of low confidence prediction data.

[0042] In the specific implementation, the offset information samples are normalized by using the min-max normalization method to map the offset distance to the [0,1] interval; the prediction data samples are normalized by Z-score to make the data scale of each detector uniform.

[0043] In one specific embodiment, the adjustment model adopts a fully connected neural network structure. The input dimension of the input layer is the sum of the offset information dimension, the predicted data dimension, and the confidence label dimension. There are three hidden layers, all using the ReLU activation function, with the number of neurons being 256, 128, and 64 respectively. A Dropout layer is added after the second hidden layer to prevent overfitting, and a Batch Normalization layer is added after the third hidden layer to accelerate training convergence. The number of neurons in the output layer is the same as the predicted data dimension, and the activation function is a linear activation function, outputting the adjusted monitoring data.

[0044] Specifically, the formula for calculating the weighted loss function L2 is as follows:

[0045] Where D is the number of the second training samples in this training, and K is the dimension of the monitoring data. For the i-th second training sample, the adjusted monitoring data is... w represents the standard data corresponding to the i-th second training sample. i Let w be the weight coefficient of the i-th second training sample; where w i =1-E i +β, E iLet be the confidence label corresponding to the i-th second training sample, with a value range of (0, 1], and β be a preset smoothing term to prevent the weight from approaching zero. In this embodiment, β = 0.1.

[0046] S303: Based on the iterative optimization results of the weighted loss function, the parameters of the preset neural network model are iteratively adjusted until the preset neural network model converges, resulting in a trained adjusted model. In specific implementation, the model parameters are updated using gradient descent based on the calculated loss value. The gradient clipping threshold can be set to 0.5. The above steps are repeated until the validation set loss value converges or the maximum number of training epochs is reached.

[0047] As mentioned above, when training and adjusting the model, confidence labels were introduced, and different weight coefficients were set for different confidence scores. This enabled the trained adjustment model to have the ability to adaptively calibrate monitoring data with low confidence, effectively suppressing the monitoring error introduced by pose changes. This provided highly reliable calibrated data for subsequent anomaly detection models, significantly reducing the misjudgment rate caused by data distortion. This ensured that power supply adjustments could be accurately made for non-anomaly detectors under low power conditions.

[0048] S4 generates model prompts based on the user's current motion status information and preset user profile features, and identifies abnormal monitoring data and corresponding abnormal detectors based on the adjusted monitoring data and the preset abnormal detection model.

[0049] Specifically, the preset user profile features are the user's pre-uploaded physical health status information and basic information to an external mobile terminal. The physical health status information includes the health status information of various parts of the body, and the basic information may include information such as gender, height, age, and historical behavior patterns.

[0050] Furthermore, the preset anomaly detection model is a rule-based expert system or a trained preset machine learning model. Specifically:

[0051] As an optional implementation, the preset anomaly detection model employs a rule-based expert system. Specifically, for detectors corresponding to different body parts, their respective normal physiological parameter ranges are pre-set. Taking an electrocardiogram (ECG) detector as an example, the normal heart rate range is 60-100 beats / minute, the normal PR interval is 120-200 milliseconds, and the normal QRS complex duration is 60-100 milliseconds. When any parameter in the adjusted monitoring data exceeds the corresponding threshold range, the data is marked as abnormal monitoring data, and the corresponding detector is marked as an abnormal detector. The threshold range can be personalized according to user profile characteristics, such as age, gender, and underlying medical history.

[0052] As an alternative implementation, the preset anomaly detection model employs a trained isolated forest model. The training process is as follows: monitoring data from several healthy subjects in resting and standard exercise states are collected to construct a training sample set; the isolated forest algorithm is used to fit the training samples, for example, setting the number of trees to 100, the subsampling size to 256, and the anomaly score threshold to 0.6. During the detection phase, the adjusted monitoring data is input into the model; when the model's output anomaly score exceeds the anomaly score threshold, it is determined to be abnormal monitoring data. This model divides the data by constructing random binary trees; abnormal data, due to significant feature differences, is more easily isolated, exhibiting better interpretability.

[0053] Specifically, the step of generating model prompts based on the current user's motion state information and preset user profile features includes the following steps:

[0054] S401, acquire user motion status information and user profile features recorded by an external mobile terminal that is connected to the wearable device.

[0055] S402 performs structured processing on user motion state information and user profile features, transforming them into natural language descriptive text, which serves as the context input for the anomaly detection model.

[0056] As mentioned above, when identifying abnormal monitoring data, based on the adjusted monitoring data, the user's current motion state information and preset user profile features are also introduced as auxiliary judgments. By integrating dynamic real-time behavior with static individual characteristics, the user's physical state information is enriched, enabling the system to more accurately predict the data value of detectors for different body parts, improving the monitoring accuracy of abnormal data, and thus enhancing the reliability of power supply distribution under different user states.

[0057] S5, when the current remaining power of the wearable device is detected to be lower than the first preset power threshold, the power supply and sampling frequency of each non-abnormal detector except the abnormal detector are adjusted until the current remaining power is restored to a level higher than the second preset power threshold; this can be understood as: the non-abnormal detector is any detector other than the abnormal detector.

[0058] Specifically, the second preset power threshold is greater than the first preset power threshold. Those skilled in the art can set the first and second preset power thresholds according to actual needs, which will not be elaborated upon here.

[0059] Specifically, adjusting the power supply and acquisition frequency of each non-abnormal detector, excluding the abnormal detector, includes the following steps:

[0060] S501, for any non-abnormal detector, when the preset priority of the non-abnormal detector is greater than the first priority threshold, the voltage and / or data acquisition frequency of the non-abnormal detector is reduced by a first preset ratio; those skilled in the art set the first priority threshold and the first preset ratio according to actual needs.

[0061] Preferably, the preset priority is either a default priority or a user-defined priority. For example, a user can set the priority of the detector according to their own physical condition or actual needs.

[0062] S502, when the preset priority of the non-abnormal detector is not greater than the first priority threshold but greater than the second priority threshold, the voltage and / or data acquisition frequency of the non-abnormal detector is reduced by a second preset ratio; wherein the second preset ratio is greater than the first preset ratio. Those skilled in the art can set the second priority threshold and the second preset ratio according to actual needs.

[0063] S503, when the preset priority of the non-abnormal detector is not greater than the second priority threshold, the power supply to the non-abnormal detector is cut off; wherein, the second priority threshold is lower than the first priority threshold.

[0064] As described above, when an anomaly detector is identified, the power supply and data acquisition frequency of the anomaly detector are maintained, while the power supply voltage and data acquisition frequency of non-anomaly detectors are appropriately reduced. Differentiated frequency reduction and voltage adjustment strategies are implemented according to the different detector priorities. This achieves a dynamic balance between precise resource allocation and core function protection. While ensuring the maximum extension of device battery life, it also ensures the core needs of user safety monitoring and significantly improves the functionality and user experience of wearable devices in low-power states.

[0065] In another embodiment, adjusting the power supply and acquisition frequency of each non-abnormal detector, excluding the abnormal detector, further includes the following steps:

[0066] S510: For any non-abnormal detector, acquire the historical power supply data and historical acquisition frequency data of the non-abnormal detector within a preset historical time period, and calculate the historical energy efficiency value of the non-abnormal detector.

[0067] Specifically, the historical energy efficiency value is the power consumed by the non-abnormal detector in a single data acquisition; it can be understood as the power consumed by the detector in a single data acquisition action.

[0068] S520, obtain the real-time monitoring data change rate of the non-anomaly detector at the current moment, and determine the data importance score of the non-anomaly detector based on the real-time monitoring data change rate; wherein, the real-time monitoring data change rate is negatively correlated with the data importance score.

[0069] In traditional methods, it is generally believed that the greater the rate of change of real-time monitoring data, the greater the fluctuation of the acquired signal, and the higher the importance of the data. However, in this embodiment, the more violent the signal fluctuations acquired by the detector, the more noise data may be generated due to factors such as loosening of the detector or body movement. In this case, the data reliability is low, and the priority of power supply and acquisition frequency should be reduced. Even if it reflects real physiological changes, that is, the current physiological state fluctuates violently, violent fluctuations are already easily identified by the anomaly detection model, and there is no need to give it high priority during the power supply adjustment phase. On the contrary, data with stable changes are often in a critical or slowly deteriorating state, and continuous monitoring is more necessary to capture anomalies. Taking the above considerations into account, the rate of change of real-time monitoring data is negatively correlated with the data importance score.

[0070] S530 constructs a two-dimensional energy consumption-importance matrix based on the historical energy efficiency value and data importance score of each non-abnormal detector, and divides the non-abnormal detectors into several adjustment levels based on the energy consumption-importance matrix. This can be understood as follows: in the constructed energy consumption-importance matrix, the historical energy efficiency value is used as the horizontal axis, and the data importance score is used as the vertical axis, mapping all detectors onto a planar matrix. Based on preset thresholds or quantiles, the matrix is ​​divided into several regions, each corresponding to an adjustment level. For example, for low energy consumption and high importance regions, the adjustment is minimal or non-adjusted; for high energy consumption and low importance regions, the frequency is significantly reduced or even cut off; for high energy consumption and high importance regions, the frequency is moderately reduced; and for low energy consumption and low importance regions, a moderate reduction is implemented.

[0071] S540 determines the voltage adjustment amount and acquisition frequency adjustment amount of the non-abnormal detector according to the divided adjustment levels and the preset level-adjustment amount mapping table. It can be understood that the level-adjustment amount mapping table stores the voltage adjustment amount and acquisition frequency adjustment amount corresponding to each adjustment level.

[0072] The above-mentioned method constructs a two-dimensional matrix using historical energy efficiency values ​​and data importance scores for hierarchical adjustment, avoiding reliance on priority or power supply adjustment based solely on energy consumption as a single indicator. Furthermore, the adjustment strategy employs hierarchical fine control, using differentiated adjustment methods for detectors in different areas, prioritizing the allocation of limited power to detectors with low energy consumption and high importance, thereby improving the economy and effectiveness of power supply allocation.

[0073] Embodiments of the present invention also provide a non-transitory computer-readable storage medium that can be disposed in an electronic device to store at least one instruction or at least one program related to implementing a method in the method embodiments, wherein the at least one instruction or the at least one program is loaded and executed by the processor to implement the power distribution method for wearable devices provided in the above embodiments.

[0074] Embodiments of the present invention also provide an electronic device, including a processor and the aforementioned non-transitory computer-readable storage medium.

[0075] While specific embodiments of the invention have been described in detail by way of example, those skilled in the art should understand that the examples are for illustrative purposes only and not intended to limit the scope of the invention. It should also be understood that various modifications can be made to the embodiments without departing from the scope and spirit of the invention. The scope of the invention is defined by the appended claims.

Claims

1. A power supply distribution method of a wearable device, the method comprising: The method includes the following steps: S1, adjust the real-time position coordinates of several detectors in the wearable device and compare them with the standard position coordinates of several pre-stored detectors to obtain offset information representing the offset state of each detector; where different detectors correspond to different body parts of the user. S2, for any detector, input the current monitoring data and corresponding offset information of the detector into the pre-trained prediction model, and output the prediction data with confidence labels; the confidence labels are negatively correlated with the offset information; S3, input the offset information, prediction data and corresponding confidence labels of each detector into the pre-trained adjustment model, calibrate the prediction data based on the confidence labels, and output the adjusted monitoring data; S4. Based on the user's current motion status information and preset user profile features, generate model prompt words, and based on the adjusted monitoring data, identify abnormal monitoring data and corresponding abnormal detectors through a preset abnormal detection model. S5, when the current remaining power of the wearable device is detected to be lower than the first preset power threshold, the power supply and sampling frequency of each non-abnormal detector except the abnormal detector are adjusted until the current remaining power is restored to a level higher than the second preset power threshold; the second preset power threshold is greater than the first preset power threshold.

2. The power distribution method for wearable devices according to claim 1, characterized in that, Step S1 includes the following steps: S101, acquire device attitude data collected by inertial measurement units located in different local areas of the wearable device; S102, based on the predefined device area model, device posture data and human kinematics model, establishes a dynamic transformation relationship from the device body coordinate system of the local area where each detector is located to the preset standard human reference coordinate system; S103, Based on the dynamic transformation relationship, the actual measured coordinates of each detector in the corresponding device body coordinate system are transformed to the standard human body reference coordinate system to obtain the normalized actual position coordinates of each detector. S104, Under the standard human body reference coordinate system, the normalized actual position coordinates of each detector are compared with the pre-stored theoretical standard position coordinates of each detector to calculate the offset information of each detector; the offset information includes at least the offset distance.

3. The power distribution method for wearable devices according to claim 1, characterized in that, The training process of the prediction model is as follows: S201, construct a first training sample set corresponding to several detectors; each training sample in the first training sample set includes an offset information sample and a corresponding original monitoring data sample, as well as a standard data sample collected by a preset reference device; S202, Construct a target machine learning model with dual output branches and a confidence mapping function; the dual output branches are used to generate prediction data and intermediate parameters representing prediction uncertainty, respectively; the confidence mapping function is used to map the output intermediate parameters into confidence scores. S203, each offset information sample and the corresponding original monitoring data sample are input into the target machine learning model, the standard data sample corresponding to each offset information sample is used as the supervision target, and the parameters of the target machine learning model are updated by optimizing the preset joint loss function to obtain the prediction model; the joint loss function is composed of a weighted average of the data reconstruction loss term and the uncertainty regularization loss term.

4. The power distribution method for wearable devices according to claim 1, characterized in that, The training process of the adjusted model is as follows: S301, Construct a second training sample set; Each training sample in the second training sample set includes an offset information sample, a corresponding standard data sample, and prediction data with confidence labels corresponding to each offset information sample output by the prediction model. S302, the offset information samples and the prediction data with confidence labels corresponding to each offset information sample output by the prediction model are used as input features, and the corresponding standard data samples are used as supervision targets to iteratively optimize the weighted loss function set in the preset neural network model; wherein, the weighted loss function represents that the weight of each training sample is negatively correlated with the confidence label; S303: Based on the iterative optimization results of the weighted loss function, the parameters of the preset neural network model are iteratively adjusted until the preset neural network model converges, thus obtaining the trained adjusted model.

5. The power distribution method for wearable devices according to claim 1, characterized in that, In step S4, generating model prompts based on the user's current motion state information and preset user profile features includes the following steps: S401, Obtain user motion status information and user profile features recorded by an external mobile terminal that is connected to the wearable device; S402 performs structured processing on user motion state information and user profile features, transforming them into natural language descriptive text, which serves as the context input for the anomaly detection model.

6. The power distribution method for wearable devices according to claim 1, characterized in that, The adjustment of the power supply and acquisition frequency for each non-abnormal detector, excluding the abnormal detector, includes the following steps: S501, for any non-abnormal detector, when the preset priority of the non-abnormal detector is greater than the first priority threshold, the voltage and / or data acquisition frequency of the non-abnormal detector is reduced by a first preset ratio; the preset priority is the default priority or the priority set by the user. S502, when the preset priority of the non-abnormal detector is not greater than the first priority threshold but greater than the second priority threshold, the voltage and / or data acquisition frequency of the non-abnormal detector is reduced by a second preset ratio; wherein, the second preset ratio is greater than the first preset ratio. S503, when the preset priority of the non-abnormal detector is not greater than the second priority threshold, the power supply to the non-abnormal detector is cut off; wherein, the second priority threshold is lower than the first priority threshold.

7. The power distribution method for wearable devices according to claim 1, characterized in that, The adjustment of the power supply and acquisition frequency for each non-abnormal detector, excluding the abnormal detector, also includes the following steps: S510, for any non-abnormal detector, acquire the historical power supply data and historical acquisition frequency data of the non-abnormal detector within a preset historical time period, and calculate the historical energy consumption efficiency value of the non-abnormal detector; the historical energy consumption efficiency value is the power consumed by the non-abnormal detector in a single data acquisition. S520, obtain the real-time monitoring data change rate of the non-anomaly detector at the current moment, and determine the data importance score of the non-anomaly detector based on the real-time monitoring data change rate; wherein, the real-time monitoring data change rate is negatively correlated with the data importance score; S530: Based on the historical energy efficiency value and data importance score of each non-anomaly detector, construct an energy consumption-importance two-dimensional matrix, and divide the non-anomaly detectors into several adjustment levels based on the energy consumption-importance two-dimensional matrix; S540 determines the voltage adjustment amount and acquisition frequency adjustment amount of the non-abnormal detector according to the divided adjustment levels and the preset level-adjustment amount mapping table.

8. The power distribution method for wearable devices according to claim 1, characterized in that, The preset anomaly detection model is a rule-based expert system or a pre-trained preset machine learning model.

9. A non-transitory computer-readable storage medium, wherein the storage medium stores at least one instruction or at least one program segment, characterized in that, The at least one instruction or the at least one program segment is loaded and executed by the processor to implement the power distribution method for the wearable device as described in any one of claims 1-8.

10. An electronic device, characterized in that, Includes a processor and the non-transitory computer-readable storage medium as described in claim 9.

Images