Low-power-consumption control method, device and equipment of sports bracelet and storage medium

Abstract

The invention relates to a low-power-consumption control method, device and equipment for a sports bracelet and a storage medium. The method comprises the following steps: constructing an initial mapping model of a motion state and chip power consumption according to first user motion data and chip power consumption data collected by a motion bracelet; inputting the second user motion data into the initial mapping model for processing to obtain a predicted power consumption value and calculating a power consumption residual error; performing priority ranking on the function tasks of the sports bracelet to obtain a task priority queue; performing frequency adjustment on the plurality of cores of the processor to obtain the working frequency of each core; performing chip voltage regulation and power consumption distribution to obtain an initial power consumption control strategy; and inputting the data of the plurality of sensors into the adaptive state observer for fusion processing, and updating to obtain a target mapping model and a target power consumption control strategy. According to the intelligent low-power-consumption control method capable of achieving self-adaptive learning and dynamic adjustment, the increasing power consumption management requirement of the sports bracelet is met.

Description

Technical Field

[0001] The present invention relates to the technical field of power consumption control, and in particular to a low power consumption control method, device, equipment and storage medium for a sports bracelet. Background Art

[0002] With the rapid development of wearable device technology, sports bracelets have attracted widespread attention as a portable health monitoring device. However, the usage scenarios of sports bracelets are complex and varied, and the user's activity status is diverse, which leads to huge challenges in power consumption management. Traditional fixed power consumption management strategies are difficult to adapt to the exercise patterns and usage habits of different users, and often have problems such as excessive power consumption or insufficient performance.

[0003] At the same time, a variety of sensors and functional modules are integrated inside the sports bracelet. How to achieve the coordinated work and efficient management of these modules under limited battery capacity has become a problem that needs to be solved urgently. Existing power consumption control methods often use simple threshold judgments or fixed scheduling strategies, which cannot fully utilize the dynamic frequency and voltage regulation capabilities of the chip, and it is difficult to accurately optimize power consumption according to the user's real-time motion status. In addition, as a long-term wearable device, the battery life of the sports bracelet directly affects the user experience. How to maximize the battery life while ensuring the normal operation of the function and adapt to the usage patterns of different users has become an important research direction for low-power control technology of sports bracelets. Summary of the invention

[0004] The main purpose of the present invention is to provide a low-power consumption control method, device, equipment and storage medium for a sports bracelet. The present invention can adaptively learn and dynamically adjust the intelligent low-power consumption control method to meet the growing power consumption management needs of sports bracelets.

[0005] To achieve the above object, the present invention provides a low power consumption control method for a sports bracelet, comprising the following steps: Performing convolutional neural network processing on the first user's motion data and chip power consumption data collected by the sports bracelet to obtain an initial mapping model between motion state and chip power consumption; Inputting the second user motion data collected in real time into the initial mapping model for processing to obtain a predicted power consumption value, and calculating a power consumption residual between the predicted power consumption value and the actual power consumption value; Prioritize the functional tasks of the sports bracelet based on the power consumption residual to obtain a task priority queue; Adjusting the frequencies of multiple cores of the processor according to the task priority queue to obtain the operating frequency of each core; Based on the operating frequency and the power consumption residual, chip voltage adjustment and power consumption allocation are performed to obtain an initial power consumption control strategy; Multiple sensor data are input into an adaptive state observer for fusion processing to obtain a system state estimate, and the initial mapping model and the initial power consumption control strategy are updated according to the system state estimate to obtain a target mapping model and a target power consumption control strategy.

[0006] The present invention also provides a low-power consumption control device for a sports bracelet, comprising: A processing module, used to perform convolutional neural network processing on the first user's motion data and chip power consumption data collected by the sports bracelet to obtain an initial mapping model between motion state and chip power consumption; A calculation module, used for inputting the second user motion data collected in real time into the initial mapping model for processing, obtaining a predicted power consumption value, and calculating a power consumption residual between the predicted power consumption value and the actual power consumption value; A sorting module, used to prioritize the functional tasks of the sports bracelet based on the power consumption residual to obtain a task priority queue; An adjustment module, used to adjust the frequencies of multiple cores of the processor according to the task priority queue to obtain the operating frequency of each core; An allocation module, configured to perform chip voltage adjustment and power consumption allocation based on the operating frequency and the power consumption residual to obtain an initial power consumption control strategy; The updating module is used to input multiple sensor data into the adaptive state observer for fusion processing to obtain a system state estimate, and update the initial mapping model and the initial power consumption control strategy according to the system state estimate to obtain a target mapping model and a target power consumption control strategy.

[0007] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0008] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned methods are implemented.

[0009] In summary, the technical solution provided by the present invention establishes a mapping model between motion state and chip power consumption through a convolutional neural network, realizes accurate modeling of the relationship between user motion mode and chip power consumption, introduces the concept of power consumption residual, and combines the task priority sorting mechanism, so that the system can dynamically adjust the task execution order according to the real-time running state, effectively improving the resource utilization efficiency. A multi-core dynamic frequency adjustment strategy is adopted to flexibly allocate processor resources for tasks of different priorities, while ensuring performance and maximizing energy consumption. Combined with chip voltage regulation and power consumption allocation technology, a more fine-grained power consumption control is achieved, effectively reducing the overall power consumption of the system. The adaptive state observer is used for multi-sensor data fusion, which improves the accuracy of system state estimation. Through online learning and parameter update mechanisms, the system can continuously adapt to user usage habits and environmental changes, and realizes continuous optimization of power consumption control strategies. The present invention organically combines multiple links such as power consumption prediction, task scheduling, frequency adjustment, and voltage regulation to build a complete low-power control system, which significantly improves the endurance and user experience of the sports bracelet. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a schematic diagram of the steps of a low power consumption control method of a sports bracelet in one embodiment of the present invention; Figure 2 It is a structural block diagram of a low power consumption control device of a sports bracelet in one embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0011] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0012] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0013] Reference Figure 1 This embodiment provides a low power consumption control method for a sports bracelet, comprising the following steps: S1, performing convolutional neural network processing on the first user motion data and chip power consumption data collected by the sports bracelet to obtain an initial mapping model between motion state and chip power consumption; Among them, the motion data of the first user is collected based on the sports bracelet, and the chip power consumption data of the sports bracelet is obtained. The motion data of the first user is preprocessed, including removing outliers, smoothing and other operations, to obtain a normalized motion data matrix to ensure the consistency and stability of the data. At the same time, the chip power consumption data is preprocessed to obtain a normalized power consumption data vector. The normalized motion data matrix is ​​input into the convolution layer of the convolutional neural network for feature extraction. In the convolution layer, the input data is slid in the local receptive field through multiple convolution kernels to capture the spatiotemporal features in the data and generate a primary feature map. In order to reduce the data dimension and reduce the computational complexity while retaining the main feature information, the primary feature map is pooled, such as maximum pooling or average pooling, to obtain a feature map after dimensionality reduction. The pooling operation reduces the spatial dimension of the feature map by downsampling on the feature map, while improving the computational efficiency and generalization ability of the model. The feature map after dimensionality reduction is input into the fully connected layer, and a nonlinear transformation is performed to extract higher-level feature information to generate a motion feature vector. The motion feature vector can effectively represent the complex features of the user's motion state. The motion feature vector is concatenated with the normalized power consumption data vector to form a comprehensive input feature vector, which contains comprehensive information about the user's motion state and chip power consumption. The input feature vector is processed by a multilayer perceptron. The multilayer perceptron is a feedforward neural network structure that processes and maps input features through multiple layers of linear and nonlinear transformations. Through the processing of the multilayer perceptron, the final predicted feature value is obtained, which represents the power consumption value predicted by the convolutional neural network for the current input data. The predicted feature value is compared with the actual power consumption value, and the mean square error between them, that is, the loss function value, is calculated. The smaller the loss function value, the smaller the gap between the power consumption value predicted by the model and the actual power consumption value, and the better the performance of the model. In order to optimize the performance of the convolutional neural network, the parameters of the convolutional neural network are back-propagated and optimized by gradient descent based on the loss function value. The back-propagation algorithm calculates the gradient of each parameter in the network according to the loss function value, and uses these gradient values ​​to adjust the weights and biases of the network so that the model can better fit the data during the next training. Using the gradient descent algorithm, the network parameters are updated at a certain learning rate to minimize the loss function value and gradually improve the prediction accuracy of the model. After multiple rounds of iteration and optimization, the updated network parameters are obtained. The updated network parameters are applied to the convolutional neural network to obtain the initial mapping model between motion state and chip power consumption.

[0014] S2, inputting the second user's motion data collected in real time into the initial mapping model for processing to obtain a predicted power consumption value, and calculating a power consumption residual between the predicted power consumption value and the actual power consumption value; Specifically, the second user's motion data is collected in real time based on the sports bracelet to reflect the various motion states of the user when wearing the sports bracelet. In order to better extract effective features, the second user's motion data is segmented by sliding windows, and the continuous time series data is cut into multiple data segments of fixed time length according to a certain time step. Each data segment contains motion state information within a period of time. The data segments are fast Fourier transformed to convert the motion data in the time domain into frequency domain feature vectors. It is easier to analyze the periodicity and vibration characteristics in the motion data in the frequency domain, so as to more effectively capture the data patterns under different motion states. The frequency domain feature vector is input into the convolution layer of the initial mapping model. The convolution layer extracts the local features of the motion data through the convolution operation and generates a preliminary feature map, which contains the primary feature information of the motion state. The preliminary feature map is subjected to the maximum pooling operation. Maximum pooling is a downsampling method that reduces the spatial dimension of the data by taking the maximum value in the local area in the feature map while retaining the most significant features. The compressed feature map is input into the fully connected layer for processing to obtain the intermediate feature vector. The fully connected layer transforms the local features extracted by the convolution layer into more abstract high-level feature representations through linear transformation and activation function, so that the model can better understand and distinguish different motion states. In order to avoid overfitting of the model, the intermediate feature vector is subjected to dropout regularization. Dropout is a regularization technology that randomly sets the output of a part of neurons to zero, thereby reducing the model's dependence on certain specific neurons and improving the generalization ability of the model. The regularized feature vector after dropout processing has stronger anti-interference ability, which helps to improve the robustness of the model under different users and different motion states. The regularized feature vector is input into the softmax classifier, which maps the input feature vector to a multi-category probability distribution, outputs the probability value of each motion state, and obtains the motion state probability distribution. Based on the motion state probability distribution and the initial mapping model, the predicted power consumption value at the current moment is calculated. The predicted power consumption value is the chip power consumption predicted by the model based on the real-time motion state. At the same time, the power consumption monitoring module built into the chip measures the current actual power consumption value in real time to obtain real-time power consumption data. The difference between the predicted power consumption value and the real-time power consumption data is calculated to obtain the power consumption residual, which reflects the deviation between the power consumption predicted by the model and the actual power consumption.

[0015] S3, prioritize the functional tasks of the sports bracelet based on the power consumption residual to obtain a task priority queue; It should be noted that the power consumption residual is normalized, and the power consumption residual data of different scales are converted to the same scale range, and the difference in dimension between different data is eliminated to obtain the normalized power consumption residual value. Based on the normalized power consumption residual value, a task priority tree is constructed to represent the priority relationship between each functional task, and the initial task tree structure is obtained. The nodes of the initial task tree structure represent each functional task, and the connection between the nodes represents the dependency relationship or priority order between the tasks. The initial task tree structure is subjected to hierarchical analysis. The hierarchical analysis method is an analysis method that hierarchically and systematically analyzes complex problems. Through hierarchical analysis, the relative importance of each functional task is compared layer by layer to determine the importance weight of each task in the overall task structure. According to the degree of influence of the task on system performance and power consumption, each task is quantitatively evaluated. The relative importance weight of each functional task is weighted and summed with the preset task basic weight to calculate the comprehensive weight value. The preset task basic weight is determined according to the historical data, importance or other specific indicators of the task, and after being combined with the relative importance weight, it reflects the actual priority of each task in the current situation. The functional tasks are preliminarily sorted based on the comprehensive weight value to generate the initial task sequence. The initial task sequence is encoded to obtain a task encoding table. The task encoding table is used to uniquely identify each functional task. According to the task encoding table, the time complexity and space complexity of each task are calculated. The time complexity measures the time resources required to execute the task, while the space complexity reflects the storage resources occupied during the execution of the task. By analyzing the time and space complexity of each task, the resource consumption of each task is quantified. The task complexity index and the comprehensive weight value are optimized for multiple objectives to balance the execution efficiency and resource consumption of the task. The time complexity, space complexity and comprehensive weight value are included in the optimization target to find an optimal weight allocation scheme to obtain the optimized task weight. Based on the optimized task weight, all functional tasks are finally sorted to form an ordered task list. The ordered task list arranges each task from high to low according to the optimized task weight, with the high priority task at the front of the list and the low priority task at the end of the list. The ordered task list is converted into a queue data structure to generate a task priority queue. The task priority queue dynamically manages the execution order of each task to ensure that in the process of power consumption control, the tasks with less impact on system performance and significant power consumption optimization effect are executed first.

[0016] S4, adjusting the frequencies of multiple cores of the processor according to the task priority queue to obtain the operating frequency of each core; Specifically, the tasks in the task priority queue are classified, and the tasks with higher priority are divided into high priority task groups, while the tasks with lower priority are divided into low priority task groups. The ratio of the number of tasks in the high priority task group and the low priority task group, that is, the task distribution ratio, is calculated. The task distribution ratio can reflect the relative proportion of high priority and low priority tasks in the current task queue. Based on the task distribution ratio, the multiple cores of the processor are classified, and the cores with stronger processing power and more resource occupation are divided into the main core group for processing tasks in the high priority task group; and the remaining cores are divided into the auxiliary core group for executing tasks in the low priority task group. The computational complexity analysis is performed on each task in the high priority task group to evaluate the amount of computing resources required for each task during execution, such as the number of floating point operations, the number of memory accesses, etc. Through the computational complexity analysis, the task computation load value of each task is obtained, and the task computation load value can quantify the demand of each task for the processor computing power. According to the task computation load value, the frequency of the main core group is preliminarily set to obtain the initial frequency of the main core. The principle of setting the initial frequency is to ensure that high-priority tasks can be completed in a shorter time and reduce the response delay caused by long task execution time. At the same time, a resource demand analysis is performed on each task in the low-priority task group. The resource demand analysis mainly evaluates the system resource requirements of each task during execution, such as memory, I / O bandwidth, etc. Through resource demand analysis, the task resource demand value of each task is obtained, which reflects the occupation of auxiliary core resources during the execution of the task. Based on the task resource demand value, the frequency of the auxiliary core group is preliminarily set to obtain the initial frequency of the auxiliary core. The initial frequency of the auxiliary core should be able to meet the resource requirements of low-priority tasks to avoid system performance degradation caused by excessive waste or insufficient resources. The initial frequency of the main core and the initial frequency of the auxiliary core are input into the dynamic voltage and frequency adjustment algorithm, and the voltage and frequency of the processor are dynamically adjusted to achieve a balance between performance and power consumption. The dynamic voltage and frequency adjustment algorithm optimizes and adjusts the initial frequency scheme according to the current system's task load, temperature, power consumption, etc., to obtain the optimized frequency scheme. The optimized frequency scheme is checked for temperature and power consumption constraints. The temperature constraint check is to prevent the processor from overheating due to excessive frequency, which may cause system stability problems; the power consumption constraint check is to prevent the overall power consumption from exceeding the design range due to excessive frequency, which may affect the battery life of the sports bracelet. After the check, the frequency value that meets the constraint conditions is obtained, and the frequency value is assigned to the corresponding processor core to determine the final operating frequency for each core.

[0017] S5, performing chip voltage adjustment and power consumption allocation based on the operating frequency and power consumption residual to obtain an initial power consumption control strategy; Among them, the operating frequency of each processor core is piecewise linear interpolated to construct a corresponding relationship curve between frequency and voltage. Piecewise linear interpolation is a method of mapping different operating frequencies to corresponding voltage values ​​by piecewise linear fitting of known data points. According to the operating frequency of the processor core, the initial power supply voltage of each core is calculated to obtain an initial voltage value set. The initial voltage value set is input into the voltage regulation module to perform voltage boundary check to ensure that the voltage value of each core is within the safe working range to avoid the situation where the voltage is too high to cause chip overheating or the voltage is too low to cause system instability. After the check, a valid voltage value set is obtained. At the same time, the power consumption residual is subjected to exponential smoothing to effectively filter out the random fluctuation components in the power consumption residual and extract the overall trend change. Through exponential smoothing, the smoothed power consumption residual trend is obtained. The power consumption residual trend can reflect the power consumption change law of the system over a period of time. Based on the smoothed power consumption residual trend, a power consumption prediction model is constructed. The power consumption prediction model uses historical power consumption data and current power consumption trend to predict the power consumption in the future time window and obtain the power consumption prediction value of the future time window. The power consumption prediction value of the future time window is compared with the preset power consumption threshold. The power consumption threshold is the upper limit of the system's power consumption in different scenarios, and is used to evaluate whether the current power consumption level exceeds expectations. By comparison, the power consumption adjustment coefficient is obtained. The power consumption adjustment coefficient reflects the degree of deviation of the current power consumption level from the power consumption threshold. The larger the coefficient, the closer the current power consumption level is to or exceeds the threshold, and stricter power consumption control is required. According to the power consumption adjustment coefficient, the effective voltage value set is fine-tuned to obtain the final voltage value set. Based on the final voltage value set, the theoretical power consumption of each core is calculated to obtain the core power consumption distribution, which reflects the power consumption of each core under the current voltage setting. The core power consumption distribution is balanced based on the improved level set algorithm. The level set algorithm is a method of iterative update to achieve balanced distribution of power consumption of different cores under the constraints of the objective function. Through the improved level set algorithm, the power consumption distribution of each core is balanced as much as possible under the premise of meeting the performance requirements, avoiding the situation where some cores are overloaded while other core resources are idle. After balancing, a balanced power consumption allocation scheme is obtained. The balanced power consumption allocation scheme is combined with the final voltage value set to obtain the initial power consumption control strategy.

[0018] S6, inputting multiple sensor data into the adaptive state observer for fusion processing to obtain a system state estimate, and updating the initial mapping model and the initial power consumption control strategy according to the system state estimate to obtain a target mapping model and a target power consumption control strategy.

[0019] Among them, time synchronization and data alignment are performed on multiple sensor data. Time synchronization is to ensure that the data of different sensors at the same time point can accurately correspond to each other to avoid data distortion caused by time differences. Data alignment is to align different types of sensor data on the time axis to form a unified, multi-dimensional aligned multi-source data matrix. The aligned multi-source data matrix is ​​input into the adaptive state observer, and the multi-sensor data fusion capability of the observer is used to preliminarily process the input data to obtain a preliminarily fused state vector. Based on the preliminarily fused state vector, a nonlinear state space model is constructed. The state space model describes the dynamic behavior and state transition relationship of the system, and can effectively capture the complex dynamic changes of the sports bracelet under different states. The untraceable transformation is applied to the system dynamic equation to generate a target sampling point set. The untraceable transformation can better estimate the state of the nonlinear system by selecting a set of special sampling points in the state space to approximate the distribution after nonlinear transformation. The generated target sampling point set is input into the state prediction function. The state prediction function predicts the system state at the future moment according to the state space model of the system to obtain the prior state estimate. The prior state estimate is a preliminary prediction of the future system state, reflecting the possible state change trend of the system under current conditions. Based on the prior state estimate, combined with the observed data, the measured prediction value is calculated and the target sequence is obtained. The target sequence represents the possible observation value sequence of the system in the future under the current state. The Kalman gain matrix is ​​updated using the target sequence. The Kalman gain matrix is ​​the weight coefficient used to adjust the state prediction value in the Kalman filter, reflecting the confidence distribution between the observed value and the predicted value of the system. By updating the Kalman gain matrix, the influence weights of different data sources in the state estimation are adaptively adjusted to obtain the adaptive weight coefficient. According to the adaptive weight coefficient, the prior state estimate is corrected to eliminate the errors caused by observation noise and model uncertainty, and the final system state estimate is obtained. Based on the system state estimate, the initial mapping model is online learned and parameter updated. Online learning continuously introduces new state information and adjusts the parameters of the model, so that the model can adaptively learn new patterns over time to obtain the target mapping model. At the same time, according to the target mapping model and the system state estimate, the initial power consumption control strategy is dynamically adjusted, and the fine control of the system power consumption is achieved by adjusting parameters such as voltage and frequency, and the target power consumption control strategy is obtained.

[0020] In one example, a convolutional neural network is used to process the first user's motion data and chip power consumption data collected by a sports bracelet to obtain an initial mapping model between motion state and chip power consumption, including: Collecting the first user's exercise data based on the sports bracelet, and obtaining the chip power consumption data of the sports bracelet; Preprocessing the first user's motion data to obtain a normalized motion data matrix, and preprocessing the chip power consumption data to obtain a normalized power consumption data vector; The normalized motion data matrix is ​​input into the convolution layer of the convolutional neural network for feature extraction to obtain a primary feature map, and the primary feature map is pooled to obtain a feature map after dimensionality reduction; The feature map after dimensionality reduction is input into the fully connected layer for nonlinear transformation to obtain a motion feature vector, and the motion feature vector is concatenated with the normalized power consumption data vector to obtain the input feature; Perform multi-layer perceptron processing on the input features to obtain predicted feature values, and calculate the mean square error between the predicted feature values ​​and the actual power consumption values ​​to obtain the loss function value; Based on the loss function value, the parameters of the convolutional neural network are back-propagated and optimized by gradient descent to obtain updated network parameters, and the updated network parameters are applied to the convolutional neural network to obtain an initial mapping model between motion state and chip power consumption.

[0021] In this example, the motion data and chip power consumption data of the first user are obtained from the sports bracelet. The motion data reflects the user's motion mode and state, while the chip power consumption data indicates the energy consumption of the bracelet during operation. The first user's motion data is cleaned and normalized. Data cleaning includes removing outliers and filling missing values ​​to ensure the integrity and consistency of the data. Normalization is to standardize the motion data to the same range (such as 0 to 1) so that data of different dimensions are comparable. The processed motion data is organized into a matrix form to obtain a motion data matrix. Assume that the motion data has time steps and features, the motion data matrix is ​​expressed as ,in Indicated in In the time step, Similarly, the chip power consumption data is normalized to obtain the normalized power consumption data vector ,in Indicates The power consumption value at each time step. The normalized motion data matrix Input the convolutional layer of the convolutional neural network for feature extraction. The convolutional layer uses multiple convolution kernels to perform convolution operations on the input data in the local receptive field, which can extract the spatiotemporal features of the data. Assume that the size of the convolution kernel is , that is, scanning at each time step time segments, the convolution operation is expressed as: ; in, Represents the value of the primary feature map obtained by the convolution operation, is the weight matrix of the convolution kernel, is the bias term. Through convolution operations, high-level motion features are automatically extracted from motion data, such as specific motion patterns or user acceleration and deceleration behaviors. Perform a pooling operation, using maximum pooling or average pooling. The pooling operation reduces the dimension of the data by downsampling, retaining the most important features while removing redundant information to obtain a feature map after dimensionality reduction. The feature map after dimensionality reduction is input into the fully connected layer for nonlinear transformation. The fully connected layer processes the input features through linear transformation and activation function to generate a motion feature vector. Assume that the feature map after dimensionality reduction is , the output of the fully connected layer is expressed as: ; in, is the weight matrix of the fully connected layer, is the bias term, is a nonlinear activation function (such as ReLU or Sigmoid), is the generated motion feature vector. With the normalized power consumption data vector Splice to get comprehensive input features ,in Represents a comprehensive feature that combines motion features and power consumption information. The input is processed by a multilayer perceptron. The multilayer perceptron consists of multiple fully connected layers, and each layer processes the input features through linear transformation and activation function. The final output is the predicted power consumption feature value. , the eigenvalue represents the model's prediction of the chip power consumption under the current input motion state. Calculate the mean square error between the predicted eigenvalue and the actual power consumption value to obtain the loss function value. The calculation formula for the mean square error is: ; in, is the loss function value, is the actual power consumption value, is the predicted power consumption value. The smaller the loss function value, the smaller the error between the power consumption predicted by the model and the actual power consumption. The parameters of the convolutional neural network are back-propagated and optimized by gradient descent. During the back-propagation process, the gradient of the loss function to each layer of the network parameters is calculated by the chain rule, and the network parameters are updated using the gradient descent algorithm. The gradient descent algorithm adjusts the weights and bias terms in the network according to the gradient of the loss function to minimize the loss function. The formula for weight update is: ; in, is the learning rate, which controls the step size of each update; is the loss function over weights After multiple rounds of iterative training, the parameters of the convolutional neural network are continuously updated, and the prediction accuracy of the model is gradually improved. The trained convolutional neural network can accurately map the motion state to the chip power consumption and obtain the initial mapping model.

[0022] In one example, the second user motion data collected in real time is input into the initial mapping model for processing to obtain a predicted power consumption value, and a power consumption residual between the predicted power consumption value and the actual power consumption value is calculated, including: The sports data of the second user is collected in real time based on the sports bracelet, and the sports data of the second user is segmented by a sliding window to obtain data segments of fixed time length; Perform fast Fourier transform on the data segment to obtain the frequency domain feature vector, and input the frequency domain feature vector into the convolution layer of the initial mapping model to obtain the preliminary feature map; Perform the maximum pooling operation on the preliminary feature map to obtain a compressed feature map, and input the compressed feature map into the fully connected layer to obtain an intermediate feature vector; Perform dropout regularization on the intermediate feature vector to obtain a regularized feature vector, and input the regularized feature vector into the softmax classifier to obtain the motion state probability distribution; Based on the motion state probability distribution and the initial mapping model, the predicted power consumption value is calculated, and the current actual power consumption value is measured through the chip's built-in power consumption monitoring module to obtain real-time power consumption data; The difference between the predicted power consumption value and the real-time power consumption data is calculated to obtain the power consumption residual.

[0023] In this example, the user motion data collected in real time is preprocessed. The sports bracelet has a variety of built-in sensors, such as accelerometers, gyroscopes, heart rate sensors, etc. These sensors can collect the user's motion status information at different time points. The second user's motion data is segmented into multiple fixed-length time segments by sliding windows. The sliding window divides the time series data into a certain time length (called the window size), and each time the sliding window moves forward, there is a certain amount of overlap (called the step size). Assume that the window size is (e.g. 2 seconds), the step size is (such as 1 second), each data segment contains The motion data within seconds is divided into multiple overlapping data segments of fixed time length. Each segment is represented by ,in, Indicates The first of the data segments The segmented data segments are fast Fourier transformed to convert the motion data in the time domain into the frequency domain. The time series signal is decomposed into a series of sine waves of different frequencies, thereby representing the frequency characteristics of the signal in the frequency domain. For each data segment , its frequency domain eigenvector is expressed as FFT ,in , each Representative The data segment is in By converting time domain data into frequency domain features, more stable motion patterns such as step frequency and breathing frequency can be extracted. These features can better reflect the user's motion state. The extracted frequency domain feature vector The input is processed by the convolution layer of the initial mapping model. The convolution layer extracts local features of the input data through multiple convolution kernels to capture local correlations in the frequency domain signal. The convolution operation is expressed as: ; in, Indicates The first segment in the convolutional layer A preliminary feature map, is the frequency domain eigenvector Quantity, is the weight of the convolution kernel, The convolution operation can perform weighted summation on different frequency components and extract the local correlation between frequency domain features. Perform a maximum pooling operation to reduce the spatial dimension of the feature map by taking the maximum value of the local area and retain important local features. Assume that the pooling window size is , then the feature map after maximum pooling is expressed as: ; in, For the The data fragments are pooled Compressed feature maps. The maximum pooling operation can effectively reduce the dimension of the feature map, improve the computational efficiency of the model, and retain the most significant feature information. The compressed feature map is input into the fully connected layer, and the intermediate feature vector is obtained after nonlinear transformation. The fully connected layer linearly transforms the input features through the weight matrix, and converts the local features into global feature representations through nonlinear mapping of the activation function. In order to prevent the model from overfitting, the intermediate feature vector is subjected to dropout regularization. Dropout is a technique that randomly sets the output of a part of neurons to zero, reduces the model's dependence on specific features, and improves the model's generalization ability. After Dropout processing, a regularized feature vector is obtained. The regularized feature vector is input into the softmax classifier, and the softmax classifier maps the input feature vector to a probability distribution of different motion states. The output of the softmax classifier is expressed as: ; in, Represents the input feature vector Classified as The probability of a class (i.e. a certain state of motion), and are the weights and biases of the softmax classifier, is the total number of categories. Through the softmax classifier, the probability distribution of different motion states, such as walking, running, and stillness, is obtained. Based on the probability distribution of motion states and the initial mapping model, the predicted power consumption value in the current motion state is calculated. Assume that the model outputs the power consumption value of the current motion state through regression , then the predicted power consumption value is expressed as: ; in, is the predicted power consumption value, is the regression function of the initial mapping model, is the motion state feature vector, It is a model parameter. The actual power consumption value is measured in real time through the power consumption monitoring module built into the chip. , get the real-time power consumption data of the sports bracelet in the current state. With real-time power consumption data Perform difference calculation to get the power consumption residual: ; in, is the power consumption residual, which indicates the deviation between the predicted power consumption and the actual power consumption. If the residual is large, it means that the model is not accurate enough in predicting the power consumption of the current state, and the model parameters or power consumption control strategy need to be adjusted.

[0024] In one example, the functional tasks of the sports bracelet are prioritized based on the power consumption residual to obtain a task priority queue, including: Normalizing the power consumption residual to obtain a normalized power consumption residual value, and constructing a task priority tree based on the normalized power consumption residual value to obtain an initial task tree structure; Perform hierarchical analysis on the initial task tree structure to obtain the relative importance weight of each functional task, and perform weighted summation of the relative importance weight and the preset task basic weight to obtain a comprehensive weight value; Preliminarily sort the functional tasks based on the comprehensive weight values ​​to obtain an initial task sequence, and encode the initial task sequence to obtain a task coding table; The time complexity and space complexity of each task are calculated according to the task coding table to obtain the task complexity index, and the task complexity index and the comprehensive weight value are optimized for multiple objectives to obtain the optimized task weight; The functional tasks are finally sorted based on the optimized task weights to obtain an ordered task list, and the ordered task list is converted into a queue data structure to obtain a task priority queue.

[0025] In this example, the power consumption residual is normalized to eliminate the dimensional differences of the power consumption residual between different tasks and time periods, and obtain the normalized power consumption residual value. A task priority tree is constructed based on the normalized power consumption residual value to obtain the initial task tree structure. The task priority tree is a hierarchical data structure that can hierarchically represent different functional tasks in the system according to their importance and priority relationships. Each node represents a functional task, and the connection relationship between nodes represents the priority dependency between tasks. The tree structure is used to describe the relative importance of tasks and their role in the overall system. A hierarchical analysis is performed on the initial task tree structure. The hierarchical analysis method compares the relative importance of different tasks pairwise to derive the relative weight of each task. The relative importance of tasks is scored and a judgment matrix is ​​constructed. Assume there are three tasks and , then the judgment matrix is ​​expressed as: ; in, Indicates the task Relative to the task The relative importance of The higher the importance, the more important it is. By calculating the eigenvector of the judgment matrix, the relative importance weight of each functional task is obtained. ,in Indicates the task The relative weight value of the task is calculated by adding the relative importance weight to the preset task basic weight to obtain the comprehensive weight value of each task. Assume that the preset task basic weight is , then the comprehensive weight value is expressed as: ; in, Indicates the task The comprehensive weight value of is the weight adjustment factor, which controls the influence ratio of relative importance weight and basic weight. , flexibly adjust task priorities under different system states. Based on the comprehensive weight value Perform a preliminary sorting of the functional tasks to obtain the initial task sequence, and encode the initial task sequence. According to the task encoding table, calculate the time complexity and space complexity of each task. The time complexity represents the time resources required for the task to execute, which is estimated by the amount of calculation and data for the task execution; while the space complexity reflects the occupation of resources such as memory during the task execution. Assume that the task The time complexity is , the space complexity is , then the task complexity index is expressed as: ; in, is the task complexity weight coefficient, which is used to balance the influence of time complexity and space complexity. Through the multi-objective optimization method, the execution time and resource consumption of each task are comprehensively considered to obtain the comprehensive complexity of the task. And the comprehensive weight value Perform multi-objective optimization to obtain the optimized task weights. The goal of multi-objective optimization is to minimize the time and resource consumption of task execution while maximizing the system importance of the task. The optimized task weights can more accurately reflect the priority of the task in the current system state. Based on the optimized task weights The functional tasks are finally sorted to obtain an ordered task list. The ordered task list is arranged from high to low according to the task priority, with high-priority tasks at the front of the list and low-priority tasks at the end. The ordered task list is converted into a queue data structure to obtain a task priority queue. The task priority queue is a first-in-first-out data structure that dynamically schedules and executes tasks in order of priority.

[0026] In one example, the frequencies of multiple cores of a processor are adjusted according to the task priority queue to obtain the operating frequency of each core, including: The task priority queue is grouped into high priority task groups and low priority task groups, and the ratio of the number of tasks in the high priority task group and the low priority task group is calculated to obtain the task distribution ratio; Based on the task distribution ratio, multiple cores of the processor are classified into a main core group and an auxiliary core group, and the computational complexity of each task in the high-priority task group is analyzed to obtain the task computation load value; The frequency of the main core group is preliminarily set according to the task calculation load value to obtain the initial frequency of the main core, and the resource requirement analysis is performed on each task in the low-priority task group to obtain the task resource requirement value; The frequency of the auxiliary core group is preliminarily set based on the task resource demand value to obtain the initial frequency of the auxiliary core, and the initial frequency of the main core and the initial frequency of the auxiliary core are input into the dynamic voltage and frequency adjustment algorithm to obtain an optimized frequency solution; The optimized frequency scheme is subjected to temperature and power consumption constraint checks to obtain frequency values ​​that meet the constraint conditions, and the frequency values ​​that meet the constraint conditions are allocated to the corresponding processor cores to obtain the operating frequency of each core.

[0027] In this example, tasks in the task priority queue are classified according to their priorities. Tasks with higher weight values ​​are classified into the high priority task group, while tasks with lower comprehensive weight values ​​are classified into the low priority task group. tasks, according to the median priority As the segmentation criteria, the priority is greater than Assign tasks to the high-priority task group, with a priority less than or equal to The number of tasks in the high-priority task group is recorded as , the number of tasks in the low priority task group is recorded as , then the task distribution ratio is expressed by the following formula: ; in, Indicates the ratio of the number of tasks in the high-priority task group to the number of tasks in the low-priority task group. When , it means that high-priority tasks in the system dominate, and more system resources are needed to ensure the timely execution of these tasks; When , it means that there are more low-priority tasks in the system, and more resources are needed to balance the processing load of these secondary tasks. The multiple cores of the processor are classified. The processor cores are divided into performance cores (primary cores) and energy efficiency cores (secondary cores). The primary core group is used to process high-priority tasks that are computationally intensive and sensitive to response time, while the secondary core group is used to process relatively simple, power-sensitive, low-priority tasks. Assume that the processor has cores, then according to the task distribution ratio To determine the number of cores in the primary and secondary core groups, use the following formula to calculate the number of primary cores: ; in, Indicates the number of cores assigned to the main core group, Indicates rounding down. The number of auxiliary cores is: ; in, Indicates the number of cores assigned to the auxiliary core group. After completing the core classification, perform computational complexity analysis on each task in the high-priority task group. Computational complexity is usually measured by the amount of computation required to execute the task, such as the number of floating-point operations, algorithm complexity, etc. Assume that the first task in the high-priority task group The computational complexity of the task is , tasks with higher computational complexity usually require higher frequency support from the main core group to ensure that the task can be completed in a shorter time. According to the computational complexity of the task, the frequency of the main core group is preliminarily set. Assume that the frequency of the main core group is , then the initial frequency of the main core is determined according to the average value of the task calculation complexity: ; in, is the initial frequency of the main core group, which indicates the base frequency that the main core group needs to allocate in order to handle the average computing load of all high-priority tasks. For each task in the low-priority task group, perform resource demand analysis. Resource demand includes memory usage, I / O bandwidth, etc., which is measured by the amount of resources required during task execution. Assume that The resource requirement of a low priority task is , which indicates the amount of system resources occupied by the task during execution. Based on the resource demand value of the low-priority task, the frequency of the auxiliary core group is initially set. Assume that the initial frequency of the auxiliary core group is , expressed as: ; in, is the initial frequency of the auxiliary core group, which indicates the base frequency that the auxiliary core group needs to allocate to handle the average resource requirements of all low-priority tasks. and auxiliary core initial frequency The dynamic voltage and frequency adjustment algorithm is used for optimization. The dynamic voltage and frequency adjustment algorithm is a regulation algorithm that can balance performance and power consumption. It controls energy consumption by dynamically adjusting the voltage and frequency of the processor. The frequency scheme optimized by the dynamic voltage and frequency adjustment algorithm is expressed as and . Perform temperature and power constraint checks on the optimized frequency solution. The temperature constraint check is to ensure that the temperature of the processor core does not exceed the safety threshold at the target frequency to avoid system instability or hardware damage caused by overheating. The power constraint check is to ensure that the optimized frequency does not cause the overall power consumption to exceed the design range, thereby affecting the system's battery life. Assuming the frequency solution and Satisfy temperature constraints and power consumption constraints , the final frequency value is expressed as: ; in, Indicates the final frequency value that satisfies all constraints. Assign the frequency value that satisfies the constraints to the corresponding processor core to obtain the operating frequency of each core.

[0028] In one example, chip voltage adjustment and power consumption allocation are performed based on the operating frequency and the power consumption residual to obtain an initial power consumption control strategy, including: Perform piecewise linear interpolation on the operating frequency to obtain a frequency-voltage correspondence curve, and calculate the initial power supply voltage of each core based on the frequency-voltage correspondence curve to obtain an initial voltage value set; The initial voltage value set is input into the voltage regulation module for voltage boundary check to obtain a valid voltage value set, and the power consumption residual is subjected to exponential smoothing to obtain a smoothed power consumption residual trend; A power consumption prediction model is constructed based on the smoothed power consumption residual trend to obtain a power consumption prediction value of a future time window, and the power consumption prediction value of the future time window is compared with a preset power consumption threshold to obtain a power consumption adjustment coefficient; The effective voltage value set is fine-tuned according to the power consumption adjustment coefficient to obtain a final voltage value set, and the theoretical power consumption of each core is calculated based on the final voltage value set to obtain the core power consumption distribution; The core power consumption distribution is balanced based on the improved level set algorithm to obtain a balanced power consumption allocation scheme, and the balanced power consumption allocation scheme is combined with the final voltage value set to obtain an initial power consumption control strategy.

[0029] In this example, the operating frequency is piecewise linearly interpolated to obtain the frequency-voltage correspondence curve. Assume that in the frequency range The relationship between frequency and voltage is linear, and piecewise linear interpolation is performed through several known key frequency points and corresponding voltage values. Assuming that the known frequency points And the corresponding voltage value , then in Frequency interval Inside, indicating frequency With voltage The linear relationship is: ; in, Indicates the operating frequency is The voltage value at and are the voltage values ​​at the end points of the interval, and is the corresponding frequency value. Through the piecewise linear interpolation method, the voltage value in each frequency interval is obtained, and a complete frequency-voltage corresponding relationship curve is constructed. After obtaining the frequency-voltage relationship curve, the initial supply voltage value of each core is calculated based on the operating frequency of different processor cores. Assume that the processor has cores, and the operating frequencies of each core are , then the initial voltage value set of each core is calculated according to the aforementioned linear interpolation formula ,in For the The initial voltage value set is input into the voltage regulation module to perform voltage boundary check to ensure that the voltage value of all cores is within a safe operating range. Usually the processor has a minimum operating voltage and maximum operating voltage If the initial voltage value of a core is If the voltage exceeds this range, it needs to be adjusted to the nearest valid voltage value. After checking the voltage boundary, the valid voltage value set is obtained. , where each voltage value All satisfied . Exponential smoothing is performed on the power consumption residual to obtain the smoothed power consumption residual trend. The power consumption residual is the difference between the actual power consumption and the predicted power consumption, which indicates the deviation of the system under the current power consumption control strategy. Exponential smoothing is used to eliminate noise and extract long-term trend changes. Assume that The power consumption residual at time , then the smoothed power consumption residual is expressed as: ; in, For the The residual value of power consumption after smoothing at each moment, is the smoothing coefficient, which is usually between 0 and 1, indicating the weight distribution of the current power consumption residual and the historical residual. Through exponential smoothing, a more stable power consumption residual trend is obtained. Based on the smoothed power consumption residual trend, a power consumption prediction model is constructed. The goal of the power consumption prediction model is to use the historical power consumption residual trend to predict the power consumption changes in the future time window, thereby providing a basis for dynamically adjusting the voltage. Assume that the input of the power consumption prediction model is the past The smoothed power consumption residual at each moment is , using methods such as linear regression models or recurrent neural networks to predict future time windows Power consumption within .in, For the The predicted power consumption value at the moment indicates the power consumption trend of the system under the current control strategy. The power consumption prediction value in the future time window is compared with the preset power consumption threshold. Compare and get the power consumption adjustment coefficient The power consumption adjustment factor indicates the risk of excessive power consumption that may occur in the system at a future time, and is expressed by the following formula: ; when When , it means that the future power consumption may exceed the threshold, and the supply voltage needs to be reduced to reduce power consumption; when When the power consumption is within the safe range, the voltage can be increased appropriately to improve performance. Fine-tune the effective voltage value set. Assume that The effective voltage of each core is , then the voltage value after fine-tuning is expressed as: ; in, For the The final voltage value of each core. After fine-tuning, the final voltage value set is obtained. , these voltage values ​​can provide the highest possible performance support while meeting the power consumption control requirements. Based on the final voltage value set, calculate the theoretical power consumption of each core. Assume that The power consumption model of each core is: ; in, For the The theoretical power consumption of each core is is the capacitive load constant of the core, is the core operating frequency, is the core supply voltage. Substituting the final voltage value set into the above formula, we get the theoretical power consumption of each core , and the core power consumption distribution of the entire system is obtained. Based on the core power consumption distribution, the improved level set algorithm is applied to balance the power consumption distribution. The level set algorithm is an algorithm for optimal segmentation. It iteratively optimizes the gradient information of the objective function so that the total power consumption of the system is evenly distributed among multiple cores. The improved level set algorithm introduces power consumption fluctuation constraints on the basis of the traditional algorithm, which can balance the power consumption load between different cores while reducing the overall power consumption. After balancing, the balanced power consumption allocation scheme is obtained. The balanced power consumption allocation scheme is combined with the final voltage value set to obtain the initial power consumption control strategy. This strategy can dynamically adjust the voltage and frequency of each core under the current system state, so that the system can reduce power consumption as much as possible while ensuring performance.

[0030] In one example, multiple sensor data are input into an adaptive state observer for fusion processing to obtain a system state estimate, and an initial mapping model and an initial power consumption control strategy are updated according to the system state estimate to obtain a target mapping model and a target power consumption control strategy, including: Perform time synchronization and data alignment on multiple sensor data to obtain an aligned multi-source data matrix, and input the aligned multi-source data matrix into an adaptive state observer to obtain a preliminary fused state vector; A nonlinear state space model is constructed based on the initially fused state vector to obtain the system dynamic equation, and an untraceable transformation is applied to the system dynamic equation to obtain a target sampling point set; Input the target sampling point set into the state prediction function to obtain a priori state estimation value, and calculate the measurement prediction value based on the priori state estimation value to obtain the target sequence; The Kalman gain matrix is ​​updated using the target sequence to obtain the adaptive weight coefficient, and the prior state estimate is corrected according to the adaptive weight coefficient to obtain the system state estimate; Based on the system state estimation, the initial mapping model is online learned and parameter updated to obtain the target mapping model, and the initial power consumption control strategy is dynamically adjusted according to the target mapping model and the system state estimation to obtain the target power consumption control strategy.

[0031] In this example, the data collected by different sensors are synchronized in time and aligned in space. Assume that the accelerometer is at time The data collected is , and the gyroscope is at time The data collected is ,when When , it is calculated by linear interpolation: ; in, is the interpolated value at time Gyroscope data. Align the sensor data spatially and transform the data of different sensors into the same reference coordinate system. Assume that the data of accelerometer and gyroscope are expressed in their local coordinate systems respectively, and align them into the global coordinate system through the rotation matrix. Assume that the rotation matrix is , then the accelerometer data and gyroscope data In the global coordinate system, it is expressed as: ; Combine multiple sensor data after time synchronization and spatial alignment into a multi-source data matrix ,in Represents the data of the heart rate sensor. The aligned multi-source data matrix Input into the adaptive state observer to obtain the initial fused state vector z The adaptive state observer can effectively suppress noise and uncertainty and extract the key state variables of the system by weighted fusion of data from different sensors. Assume that the fusion function of the observer is , then the state vector z It is expressed as: ; in, is the observer's adaptive weight vector, which improves the accuracy of state observation by dynamically adjusting these weights. The initial fused state vector Contains comprehensive features of multiple sensor information, such as movement speed, direction, and heart rate status. Based on the initial fusion state vector , construct a nonlinear state space model to describe the dynamic behavior of the system. Nonlinear state space models are usually composed of state equations and observation equations. The state equations represent the evolution of the system state over time, while the observation equations describe the relationship between the measured data and the state variables. Assume that the nonlinear state equation of the system is: ; in, is the state transfer function, is the control input of the system, is the process noise. The observation equation is: ; in, is the measured value, is the observation function, The measurement noise is used. Based on the nonlinear state space model, the untraceable transformation is used to deal with the state estimation problem of nonlinear systems. The untraceable transformation is a technique for efficiently processing nonlinear transformations. By selecting a set of uniformly distributed sampling points (called Sigma points) in the state space, the nonlinear transformation of the state variables is approximately solved. Determine the set of sampling points , and calculate its corresponding weight , and then perform nonlinear transformation of the state equation on the sampling points: ; in, Represents a nonlinear state transfer function The transformed target sampling point set. Input the target sampling point set into the state prediction function to obtain the system's prior state estimate ; in, is the prior state estimate, which indicates the predicted state of the system at the next moment under the current system state and control input. Based on the prior state estimate, the measured prediction value is calculated and compare it with the actual measured value Compare and get the target sequence ; The target sequence represents the residual between the actual measured value and the predicted value, reflecting the accuracy of the system state estimation. Update the Kalman gain matrix , the Kalman gain matrix is ​​used to feed the measurement residuals back into the state estimation, thereby adjusting the state prediction of the system. The updated Kalman gain matrix is ​​expressed as: ; in, is the covariance matrix between states and measurements, is the covariance matrix of the measured value. The size of the Kalman gain matrix determines the degree of correction of the residual feedback to the state estimate. Based on the Kalman gain matrix, the adaptive weight coefficient is obtained , determines how to fuse the information of different sensors into the state estimation. According to the adaptive weight coefficient, the prior state estimation value is corrected to obtain the system state estimation: ; Corrected system state estimate It more accurately reflects the real state of the system at the current moment and eliminates the deviation caused by measurement noise and model uncertainty. Based on the system state estimation, the initial mapping model is learned online and its parameters are updated. Online learning means that after receiving new state estimation data, the model parameters are dynamically updated to adapt to the changes in the system state. Assume that the parameters of the mapping model are , update the parameters by gradient descent: ; in, is the learning rate, is the loss function, which represents the error between the state estimate and the target value. , and obtain the target mapping model to better adapt to the current system state. According to the target mapping model and system state estimation, the initial power consumption control strategy is dynamically adjusted to obtain the target power consumption control strategy. The target power consumption control strategy can flexibly adjust the processor's voltage, frequency and other parameters according to the real-time changes of the system state, optimize the system power consumption and improve the system performance.

[0032] Reference Figure 2 This embodiment provides a low-power consumption control device for a sports bracelet, including: Processing module 1, used for performing convolutional neural network processing on the first user motion data and chip power consumption data collected by the sports bracelet to obtain an initial mapping model between motion state and chip power consumption; Calculation module 2, used for inputting the second user motion data collected in real time into the initial mapping model for processing, obtaining a predicted power consumption value, and calculating a power consumption residual between the predicted power consumption value and the actual power consumption value; A sorting module 3 is used to prioritize the functional tasks of the sports bracelet based on the power consumption residual to obtain a task priority queue; An adjustment module 4 is used to adjust the frequencies of multiple cores of the processor according to the task priority queue to obtain the operating frequency of each core; Allocation module 5, used to adjust chip voltage and allocate power consumption based on operating frequency and power consumption residual, and obtain an initial power consumption control strategy; The updating module 6 is used to input multiple sensor data into the adaptive state observer for fusion processing to obtain a system state estimate, and update the initial mapping model and the initial power consumption control strategy according to the system state estimate to obtain a target mapping model and a target power consumption control strategy.

[0033] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to the above method embodiment, which will not be repeated here.

[0034] Reference Figure 3In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0035] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0036] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0037] In summary, a low-power consumption control method for a sports bracelet is proposed in an embodiment of the present invention. A mapping model of motion state and chip power consumption is established through a convolutional neural network, which realizes accurate modeling of the relationship between user motion mode and chip power consumption, introduces the concept of power consumption residual, and combines the task priority sorting mechanism, so that the system can dynamically adjust the task execution order according to the real-time running state, effectively improving the resource utilization efficiency. A multi-core dynamic frequency adjustment strategy is adopted to flexibly allocate processor resources for tasks of different priorities, and minimize energy consumption while ensuring performance. Combined with chip voltage regulation and power consumption allocation technology, a more fine-grained power consumption control is achieved, effectively reducing the overall power consumption of the system. The adaptive state observer is used to perform multi-sensor data fusion, which improves the accuracy of system state estimation. Through online learning and parameter update mechanisms, the system can continuously adapt to the user's usage habits and environmental changes, and realizes the continuous optimization of the power consumption control strategy. The present invention organically combines multiple links such as power consumption prediction, task scheduling, frequency adjustment, and voltage regulation to construct a complete low-power consumption control system, which significantly improves the endurance and user experience of the sports bracelet.

[0038] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0039] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0040] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A low power consumption control method for a sports bracelet, characterized in that: The following steps are involved: Performing convolutional neural network processing on the first user's motion data and chip power consumption data collected by the sports bracelet to obtain an initial mapping model between motion state and chip power consumption; Inputting the second user motion data collected in real time into the initial mapping model for processing to obtain a predicted power consumption value, and calculating a power consumption residual between the predicted power consumption value and the actual power consumption value; Prioritize the functional tasks of the sports bracelet based on the power consumption residual to obtain a task priority queue; Adjusting the frequencies of multiple cores of the processor according to the task priority queue to obtain the operating frequency of each core; Based on the operating frequency and the power consumption residual, chip voltage adjustment and power consumption allocation are performed to obtain an initial power consumption control strategy; Multiple sensor data are input into an adaptive state observer for fusion processing to obtain a system state estimate, and the initial mapping model and the initial power consumption control strategy are updated according to the system state estimate to obtain a target mapping model and a target power consumption control strategy.

2. The low power consumption control method of the sports bracelet according to claim 1, characterized in that: The convolutional neural network is used to process the first user motion data and chip power consumption data collected by the sports bracelet to obtain an initial mapping model between motion state and chip power consumption, including: Collecting the first user's motion data based on the sports bracelet, and acquiring the chip power consumption data of the sports bracelet; Preprocessing the first user motion data to obtain a normalized motion data matrix, and preprocessing the chip power consumption data to obtain a normalized power consumption data vector; Inputting the normalized motion data matrix into a convolutional layer in a convolutional neural network for feature extraction to obtain a primary feature map, and performing a pooling operation on the primary feature map to obtain a feature map after dimensionality reduction; Inputting the dimension-reduced feature map into a fully connected layer for nonlinear transformation to obtain a motion feature vector, and concatenating the motion feature vector with the normalized power consumption data vector to obtain an input feature; Performing multi-layer perceptron processing on the input features to obtain predicted feature values, and calculating the mean square error between the predicted feature values ​​and the actual power consumption values ​​to obtain a loss function value; Based on the loss function value, the parameters of the convolutional neural network are back-propagated and gradient-descent optimized to obtain updated network parameters, and the updated network parameters are applied to the convolutional neural network to obtain an initial mapping model of motion state and chip power consumption.

3. The low power consumption control method of the sports bracelet according to claim 2, characterized in that: The step of inputting the real-time collected second user motion data into the initial mapping model for processing to obtain a predicted power consumption value, and calculating a power consumption residual between the predicted power consumption value and the actual power consumption value includes: The sports data of the second user is collected in real time based on the sports bracelet, and the sports data of the second user is segmented by a sliding window to obtain data segments of fixed time length; Performing a fast Fourier transform on the data segment to obtain a frequency domain feature vector, and inputting the frequency domain feature vector into a convolution layer of the initial mapping model to obtain a preliminary feature map; Performing a maximum pooling operation on the preliminary feature map to obtain a compressed feature map, and inputting the compressed feature map into a fully connected layer to obtain an intermediate feature vector; Performing dropout regularization processing on the intermediate feature vector to obtain a regularized feature vector, and inputting the regularized feature vector into a softmax classifier to obtain a motion state probability distribution; Based on the motion state probability distribution and the initial mapping model, a predicted power consumption value is calculated, and the current actual power consumption value is measured by a power consumption monitoring module built into the chip to obtain real-time power consumption data; The predicted power consumption value and the real-time power consumption data are difference calculated to obtain a power consumption residual.

4. The low power consumption control method of the sports bracelet according to claim 1, characterized in that: The prioritizing the functional tasks of the sports bracelet based on the power consumption residual to obtain a task priority queue includes: Normalizing the power consumption residual to obtain a normalized power consumption residual value, and constructing a task priority tree based on the normalized power consumption residual value to obtain an initial task tree structure; Performing a hierarchical analysis on the initial task tree structure to obtain the relative importance weight of each functional task, and performing a weighted summation of the relative importance weight and the preset task basic weight to obtain a comprehensive weight value; Preliminarily sorting the functional tasks based on the comprehensive weight values ​​to obtain an initial task sequence, and encoding the initial task sequence to obtain a task coding table; Calculating the time complexity and space complexity of each task according to the task coding table to obtain a task complexity index, and performing multi-objective optimization on the task complexity index and the comprehensive weight value to obtain an optimized task weight; The functional tasks are finally sorted based on the optimized task weights to obtain an ordered task list, and the ordered task list is converted into a queue data structure to obtain a task priority queue.

5. The low power consumption control method of the sports bracelet according to claim 1, characterized in that: The frequency adjustment of multiple cores of the processor according to the task priority queue to obtain the operating frequency of each core includes: Grouping the tasks in the task priority queue to obtain a high priority task group and a low priority task group, and calculating the ratio of the number of tasks in the high priority task group and the low priority task group to obtain a task distribution ratio; Based on the task distribution ratio, multiple cores of the processor are classified to obtain a main core group and an auxiliary core group, and computational complexity analysis is performed on each task in the high-priority task group to obtain a task computation load value; Preliminarily setting the frequency of the main core group according to the task calculation load value to obtain the main core initial frequency, and performing resource requirement analysis on each task in the low priority task group to obtain the task resource requirement value; Preliminarily setting the frequency of the auxiliary core group based on the task resource requirement value to obtain the auxiliary core initial frequency, and inputting the main core initial frequency and the auxiliary core initial frequency into a dynamic voltage frequency adjustment algorithm to obtain an optimized frequency solution; The optimized frequency scheme is subjected to temperature and power consumption constraint checks to obtain a frequency value that satisfies the constraint conditions, and the frequency value that satisfies the constraint conditions is allocated to a corresponding processor core to obtain an operating frequency of each core.

6. The low power consumption control method of the sports bracelet according to claim 1, characterized in that: The chip voltage adjustment and power consumption allocation are performed based on the operating frequency and the power consumption residual to obtain an initial power consumption control strategy, including: Performing piecewise linear interpolation on the operating frequency to obtain a frequency-voltage correspondence curve, and calculating an initial power supply voltage of each core based on the frequency-voltage correspondence curve to obtain an initial voltage value set; Inputting the initial voltage value set into a voltage regulation module for voltage boundary check to obtain a valid voltage value set, and performing exponential smoothing on the power consumption residual to obtain a smoothed power consumption residual trend; Building a power consumption prediction model based on the smoothed power consumption residual trend to obtain a power consumption prediction value for a future time window, and comparing the power consumption prediction value for the future time window with a preset power consumption threshold to obtain a power consumption adjustment coefficient; Fine-tuning the effective voltage value set according to the power consumption adjustment coefficient to obtain a final voltage value set, and calculating the theoretical power consumption of each core based on the final voltage value set to obtain a core power consumption distribution; The core power consumption distribution is balanced based on an improved level set algorithm to obtain a balanced power consumption allocation scheme, and the balanced power consumption allocation scheme is combined with the final voltage value set to obtain an initial power consumption control strategy.

7. The low power consumption control method of the sports bracelet according to claim 1, characterized in that: The step of inputting multiple sensor data into an adaptive state observer for fusion processing to obtain a system state estimate, and updating the initial mapping model and the initial power consumption control strategy according to the system state estimate to obtain a target mapping model and a target power consumption control strategy includes: Performing time synchronization and data alignment on multiple sensor data to obtain an aligned multi-source data matrix, and inputting the aligned multi-source data matrix into an adaptive state observer to obtain a preliminary fused state vector; Building a nonlinear state space model based on the initially fused state vector to obtain a system dynamic equation, and applying an untraceable transformation to the system dynamic equation to obtain a target sampling point set; Inputting the target sampling point set into a state prediction function to obtain a priori state estimation value, and calculating a measurement prediction value based on the priori state estimation value to obtain a target sequence; Using the target sequence to update the Kalman gain matrix to obtain an adaptive weight coefficient, and correcting the prior state estimate according to the adaptive weight coefficient to obtain a system state estimate; Based on the system state estimation, the initial mapping model is subjected to online learning and parameter updating to obtain a target mapping model, and the initial power consumption control strategy is dynamically adjusted according to the target mapping model and the system state estimation to obtain a target power consumption control strategy.

8. A low-power control device for a sports bracelet, characterized in that: For implementing the steps of the method according to any one of claims 1 to 7, the device comprises: A processing module, used to perform convolutional neural network processing on the first user's motion data and chip power consumption data collected by the sports bracelet to obtain an initial mapping model between motion state and chip power consumption; A calculation module, used for inputting the second user motion data collected in real time into the initial mapping model for processing, obtaining a predicted power consumption value, and calculating a power consumption residual between the predicted power consumption value and the actual power consumption value; A sorting module, used to prioritize the functional tasks of the sports bracelet based on the power consumption residual to obtain a task priority queue; An adjustment module, used to adjust the frequencies of multiple cores of the processor according to the task priority queue to obtain the operating frequency of each core; An allocation module, configured to perform chip voltage adjustment and power consumption allocation based on the operating frequency and the power consumption residual to obtain an initial power consumption control strategy; The updating module is used to input multiple sensor data into the adaptive state observer for fusion processing to obtain a system state estimate, and update the initial mapping model and the initial power consumption control strategy according to the system state estimate to obtain a target mapping model and a target power consumption control strategy.

9. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images