Neural network power consumption management method of Bluetooth earphone

Abstract

The invention relates to a neural network power consumption management method for a Bluetooth headset, in particular to a method for predicting power consumption and dynamically adjusting performance parameters of the Bluetooth headset by using a recurrent neural network, which comprises the following steps of: (1) acquiring a historical power consumption data set of the Bluetooth headset; (2) predicting current power consumption xt + T by using a recurrent neural network (RNN); (3) comparing the predicted current power consumption xt + T with the current power consumption demand xt + T '; (4) determining an adjustment strategy based on a comparison result; and (5) the CPU frequency FCPU and the wireless transmission power PTX of the Bluetooth earphone are adjusted according to the adjustment strategy, the multi-level judgment and adjustment strategy enables the system to optimize performance at different time scales, short-term power consumption fluctuation can be dealt with, and the long-term energy-saving requirement can be met.

Description

TECHNICAL FIELD

[0001] The present application relates to a neural network power management method for Bluetooth earphones, in particular a method for predicting power consumption using recurrent neural networks and dynamically adjusting performance parameters of Bluetooth earphones. BACKGROUND

[0002] With the rapid development of wireless audio technology, Bluetooth earphones have become an indispensable electronic device in people's daily life. However, with the increasing functions, the power management of Bluetooth earphones is facing greater and greater challenges. Traditional power management methods mainly rely on fixed thresholds or simple rules to adjust device performance, which often performs poorly in practical applications.

[0003] Currently, the industry generally adopts a management method based on preset power consumption thresholds. This method usually sets several fixed power consumption levels, and switches between these levels according to the current power consumption to adjust CPU frequency and wireless transmission power. Although this method is simple to implement, it has many problems. First, fixed thresholds cannot adapt to different users' usage habits and changing usage environments. Second, it can only passively respond to current power consumption conditions, lacks foresight, and often leads to poor balance between performance and power consumption. Third, this method is difficult to cope with complex usage scenarios, such as large power consumption fluctuations and unstable performance when multitasking or high-intensity audio processing.

[0004] Another common method is dynamic adjustment based on battery power. This method adjusts device performance according to the remaining power, but it ignores users' actual needs and usage patterns, which may excessively reduce performance when the power is low, affecting user experience. In addition, this method also cannot effectively predict and manage short-term power consumption peaks, easily leading to performance degradation or unexpected shutdown of the device under high load conditions.

[0005] Some manufacturers have tried to use machine learning methods to optimize power management, but these methods mostly stay on offline training and simple linear prediction models, making it difficult to capture the complex nonlinear characteristics of power consumption changes and adapt to changes in user behavior in real time.

[0006] The limitations of these existing technical solutions have led to a series of problems: unstable battery life, large fluctuations in audio performance, poor user experience, and even potential impact on device lifespan. Therefore, there is an urgent need for a power management method that can intelligently predict power consumption, dynamically adjust performance, and adapt to different users and scenarios. SUMMARY

[0007] The present application is an innovative solution to the above problems. It uses a power consumption prediction model based on recurrent neural networks, combined with multi-level judgment and adjustment strategies, to achieve intelligent and fine-grained management of Bluetooth headset power consumption. This method not only accurately predicts future power consumption needs, but also dynamically adjusts CPU frequency and wireless transmission power based on prediction results, maximizing battery life while ensuring performance.

[0008] The present application proposes a neural network power management method for Bluetooth headset, including the following steps:

[0009] (1) Obtain the historical power consumption data set X = {x t ,x t+1 ,…,x t+T-1} of the Bluetooth headset, where x t is the power consumption at time t, and T is the prediction time window;

[0010] (2) Use recurrent neural network RNN to predict the current power consumption x t+T , the prediction model of RNN is as follows: h t =

[0011] tanh(W hx x t +W hh h t-1 +b h )

[0012] y t =W hy h t +b y

[0013] x t+T =σ(y t )

[0014] Where h t is the hidden state at time t, W hx is the weight matrix of the hidden layer, W hh is the weight matrix of the hidden layer to the hidden layer, W hy is the weight matrix of the hidden layer to the output layer, b h and b u are the bias vectors of the hidden layer and the output layer respectively, and σ is the sigmoid activation function.

[0015] (3) Compare the predicted current power consumption x t +T with the current power consumption demand x t +T′;

[0016] (4) Determine the adjustment strategy based on the comparison result;

[0017] (5) adjusting the CPU frequency F of the Bluetooth headset according to the adjustment strategy CPU and the wireless transmission power P TX .

[0018] Preferably, the step (4) of determining the adjustment strategy based on the comparison result comprises: if x t+T < x′ t+T , increasing the wireless transmission power P TX and the CPU frequency F CPU .

[0019] Otherwise, determining whether the following condition is satisfied:

[0020] (x t +x t+1 +…+x t+t-1 )>μx t+T

[0021] wherein μ is a long-term power consumption evaluation coefficient, 0<μ<1.

[0022] Preferably, when (x t +x t+1 +…+x t+T-1 )>μx t+T is satisfied, the method further comprises the following steps:

[0023] (a) calculating the average power consumption y t in the time period: wherein y=1,2,…,T-1;

[0024] (b) comparing the obtained T-1 average power consumptions y t with the power consumption demand y′ t in the corresponding time period respectively;

[0025] (c) if there exists y t <y′ t , adjusting the current power consumption higher than the power consumption demand to the corresponding expected current power consumption x″ t+T , and increasing the wireless transmission power P TX and the CPU frequency F CPU .

[0026] (d) if there is no y t <y′ t , determining the adjustment strategy according to x′ t+T calculated from the current power consumption x t+T and the corresponding power consumption demand.

[0027] Preferably, the method of adjusting the current power consumption in the step (c) is:

[0028] x″t+T =x′ t+T +α*(x t+T -x′ t+T )

[0029] Where α is the adjustment coefficient, 0 < α < 1, used to control the adjustment range.

[0030] Preferably, the method for adjusting the CPU frequency FCPU and the wireless transmission power PTX in step (5) is as follows:

[0031] F CPU_new =F CPU +a1

[0032] P TX_new =P TX +b1

[0033] in,

[0034]

[0035] a max and b max These represent the maximum adjustable ranges of CPU frequency and wireless transmission power, respectively, with β and γ being adjustment coefficients, and P... c and P w These are the power consumption coefficients for the CPU and the wireless transmission module, respectively.

[0036] Preferably, the method further includes the following steps:

[0037] (6) Calculate performance P, delay L and battery life B as evaluation indicators;

[0038] (7) Determine whether any evaluation index exceeds the corresponding preset threshold. If not, end; if yes, proceed to step (8).

[0039] (8) Determine the weighted sum S and weighted difference D of performance, delay, and battery life:

[0040] S = w1P + w2L + w3B

[0041] D=|w1P-w2L|+|w2L-w3B|+|w3B-w1P|

[0042] Where P, L, and B represent the normalized values ​​of performance, delay, and battery life, respectively, and w1, w2, and w3 are weighting coefficients.

[0043] (9) Set the expected weighted sum S′ and the expected weighted difference D′ for each evaluation indicator; (10) Determine the change amount corresponding to each evaluation indicator based on S, S′, D and D′.

[0044] Preferably, the method of determining the power consumption management strategy in step (10) comprises:

[0045] If S < S', increase the wireless transmission power PTX and the CPU frequency FCPU;

[0046] If D > D', increase the wireless transmission power PTX and the CPU frequency FCPU;

[0047] If D < D', keep the wireless transmission power PTX at the current value and keep the CPU frequency FCPU higher than the current value.

[0048] Preferably, the method of determining the change amount corresponding to each evaluation index in step (10) comprises:

[0049] F CPU-new = F CPU

[0050] F CPU_new = F CPU + a1

[0051]

[0052] If the corresponding evaluation index is the CPU frequency, then: If the corresponding evaluation index is the wireless transmission power, then:

[0053] P TX_new = P TX + b1

[0054]

[0055] If the corresponding evaluation index is the audio processing complexity C Audio , then:

[0056]

[0057] where a max , b max and c max are the maximum allowed adjustment amplitudes of the CPU frequency, the wireless transmission power and the audio processing complexity respectively, η, ρ and ω are adjustment coefficients, F' CPU , P' TX and C' Audio are the corresponding target values.

[0058] Preferably, the method further comprises the following steps:

[0059] (11) determining whether the following condition is met:

[0060] max(x t , x t+1 , …, x t+T-1(1 + θ) * x t+T

[0061] where θ is a power fluctuation tolerance coefficient, θ > 0;

[0062] (12) If the condition is met, perform smoothing processing:

[0063] x t+i_smooth = λx t+i + (1 - λ) x t+T

[0064] where i = 0, 1, …, T-1, λ is a smoothing coefficient, 0 < λ < 1;

[0065] (13) Re-execute steps (2) to (5) using the smoothed power consumption data.

[0066] Preferably, the following steps are also included:

[0067] (14) Update the RNN model parameters regularly, and the update method is:

[0068] Use the actual power consumption data and the predicted power consumption data of the last N time windows to construct a training set;

[0069] Use the back propagation algorithm to minimize the loss function L:

[0070]

[0071] where x i + T is the actual power consumption, is the predicted power consumption;

[0072] Update the weight matrix W h x, W hh , W hy and the bias vector b h , b y :

[0073]

[0074] where η is the learning rate, W represents any weight matrix, and b represents any bias vector.

[0075] The beneficial effects of the present application mainly include the following aspects:

[0076] The core of the present application is to solve the problems of inaccurate power consumption prediction, inflexible adjustment strategy, and inability to adapt to complex scenarios in the prior art. By introducing deep learning technology, the present application can capture the complex nonlinear characteristics of power consumption changes and achieve more accurate prediction. At the same time, the multi-level judgment and adjustment strategy enables the system to optimize performance on different time scales, both coping with short-term power fluctuations and meeting long-term energy-saving needs.

[0077] More importantly, the present application ingeniously solves the long-standing algorithm mechanism contradiction in power consumption management. In traditional methods, improving prediction accuracy often means more complex models and higher computational overhead, which contradicts the goal of reducing power consumption. The present application designs a lightweight RNN model and an efficient parameter update mechanism to control the additional computational overhead to a minimum range while ensuring prediction accuracy. This design not only solves the contradiction between accuracy and efficiency, but also achieves synergistic effect of the two.

[0078] In addition, the present application achieves complementary and additive effects of optimization among multiple performance indicators. For example, by accurately controlling CPU frequency and wireless transmission power, not only the overall power consumption is reduced, but also the audio delay is reduced and the transmission stability is improved. These improvements promote each other, ultimately achieving comprehensive improvement of battery life, audio performance, and user experience.

[0079] In summary, the Bluetooth headset neural network power consumption management method proposed by the present application effectively solves many problems in the prior art through innovative algorithm design and strategy optimization. It not only significantly improves the energy efficiency and performance stability of the device, but also opens up a new research direction for power consumption management of intelligent audio devices, with important theoretical value and broad application prospects. BRIEF DESCRIPTION OF DRAWINGS

[0080] Figure 1 The logic diagram for RNN prediction of power consumption of the present application.

[0081] Figure 2 The logic diagram for long-term power consumption evaluation of the present application.

[0082] Figure 3 The logic diagram for average power consumption calculation and comparison of the present application.

[0083] Figure 4 The logic diagram for adjustment strategy determination of the present application.

[0084] Figure 5 The logic diagram for execution of adjustment of the present application.

[0085] Figure 6 The logic diagram for evaluation index calculation of the present application.

[0086] Figure 7 Logical block diagram for updating the RNN model of the present application. DETAILED DESCRIPTION

[0087] In order to further explain the technical means and effects taken by the present application to achieve the predetermined object of the application, the specific embodiments, structures, features and effects thereof are described in detail below in combination with the drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0088] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0089] Embodiment 1

[0090] Referring to Figures 1-7 , the application relates to a neural network power consumption management method for a Bluetooth headset, comprising the following steps:

[0091] (1) obtaining a historical power consumption data set X = {x t ,x t+1 ,…,x t+T-1} of the Bluetooth headset, wherein x t is the power consumption at time t, and T is a prediction time window;

[0092] (2) predicting the current power consumption x t+T using a recurrent neural network RNN, and the prediction model of the RNN is as follows: h t = tanh(W hx x t +W hh h t-1 +b h )

[0093] y t = W hy h t +b y

[0094] x t+T = σ(y t )

[0095] wherein h t is the hidden state at time t, W hx is the weight matrix of the input to the hidden layer, W hh is the weight matrix of the hidden layer to the hidden layer, W hy is the weight matrix of the hidden layer to the output layer, b h and b ubias vectors of hidden layer and output layer respectively, and σ is a sigmoid activation function;

[0096] (3) comparing the predicted current power consumption x t +T with the actual current power consumption demand x t +T′;

[0097] (4) determining an adjustment strategy based on the comparison result;

[0098] (5) adjusting the CPU frequency F CPU and the wireless transmission power P TX of the Bluetooth headset according to the adjustment strategy.

[0099] Specifically, when collecting the historical power consumption dataset, we consider a prediction time window T. For example, if T is set to 10, we will collect the power consumption data of the last 10 time points. The advantage of this is that it can capture both short-term power fluctuations and not increase the computational burden due to excessive data volume.

[0100] Next, the present application uses a recurrent neural network (RNN) to predict the current power consumption. The advantage of RNN is that it can handle time series data, which is very suitable for our scenario that needs to consider the historical power consumption changes.

[0101] Then, the present application compares the predicted current power consumption with the actual power consumption demand. This step is the key to the whole method, because it determines whether the performance parameters of the Bluetooth headset need to be adjusted. The comparison result directly affects the subsequent adjustment strategy.

[0102] Preferably, the step (4) of determining an adjustment strategy based on the comparison result comprises: if x t+T <x′ t+T , then increasing the wireless transmission power P T X and the CPU frequency F C PU.

[0103] Otherwise, it is determined whether the following condition is met:

[0104] (x t +x t+1 +…+x t+t-1 )>μx t+T

[0105] wherein μ is a long-term power consumption evaluation coefficient, 0<μ<1.

[0106] Preferably, when (x t +x t+1 +…+x t+T-1 )>μx t+T is met, the following steps are further included:

[0107] (a) Calculate the average power consumption y over a time period t : where y = 1, 2, …, T-1;

[0108] (b) Compare the obtained T-1 average power consumptions y t with the corresponding power consumption demands y′ t in the time period respectively;

[0109] (c) If there exists y t <y′ t , adjust the current power consumption that is higher than the power consumption demand to a corresponding expected current power consumption x″ t+T , and increase the wireless transmission power P TX and CPU frequency F CPU ;

[0110] (d) If there does not exist y t <y′ t , determine the adjustment strategy according to the current power consumption x t+T and the corresponding power consumption demand x′ t+T .

[0111] Preferably, the present application adopts a multi-level judgment method when determining the adjustment strategy. First, if the predicted power consumption is lower than the actual demand, we will increase the wireless transmission power and CPU frequency to meet the performance demand.

[0112] Secondly, if the predicted power consumption is higher than the actual demand, we will further judge whether the following conditions are met:

[0113] where μ is a long-term power consumption evaluation coefficient, taking a value in the range of 0 to 1. For example, we can set μ to 0.9. This means that if the total power consumption in the prediction time window exceeds 90% of the predicted power consumption, we consider that there may be long-term power consumption pressure and need to make more detailed adjustments.

[0114] In an embodiment of the present application, if the above conditions are met, we will further calculate the average power consumption in the time period and compare it with the power consumption demand in the corresponding time period. This approach can more accurately locate the source of power consumption pressure, so as to make more targeted adjustments.

[0115] where y can take values from 1 to T-1. For example, if T is 10, we can calculate the average power consumption in 9 different time periods.

[0116] Next, the present application proposes an intelligent adjustment method. If the average power consumption in a certain time period is lower than the corresponding power consumption demand, we will adjust the current power consumption that is higher than the power consumption demand to an expected value. The calculation method of this expected value is:

[0117] x″ t+T = x′ t+T + α * (x t+T - x′ t+T )

[0118] where α is an adjustment coefficient, 0 < α < 1, used to control the amplitude of adjustment. In practical applications, we can set the value of α according to the specific situation of the Bluetooth headset. For example, if we want to adjust more gently, we can set α to 0.3; if we want to adjust more aggressively, we can set α to 0.7.

[0119] Another innovation of the present application is the fine adjustment of CPU frequency and wireless transmission power. Specifically, the method for adjusting CPU frequency FCPU and wireless transmission power PTX in step (5) is:

[0120] F CPU_new = F CPU + a1

[0121] P TX_new = P TX + b1

[0122] where,

[0123]

[0124] a max and b max are the maximum adjustable amplitudes of CPU frequency and wireless transmission power, β and γ are adjustment coefficients, P c and P w are the power consumption coefficients of CPU and wireless transmission module. The advantage of this adjustment method is that it takes into account the actual power consumption gap and sets an upper limit for adjustment, avoiding system instability caused by excessive adjustment.

[0125] In another embodiment of the present application, we introduce performance, delay and battery life as evaluation indicators. The selection of these three indicators is based on the aspects that Bluetooth headset users care most about. We first determine whether these indicators exceed the preset threshold, and if they do, more complex adjustments are needed.

[0126] Specifically, it further includes the following steps:

[0127] (6) Calculate performance P, delay L and battery life B as evaluation indicators;

[0128] (7) Determine whether any of the evaluation indicators exceeds the corresponding preset threshold, if not, end; if yes, execute step (8);

[0129] (8) the weighted sum S and the weighted difference D of performance, delay and battery life are determined:

[0130] S = w1P + w2L + w3B

[0131] D = |w1P - w2L| + |w2L - w3B| + |w3B - w1P|

[0132] where P, L, B represent the normalized values of performance, delay and battery life respectively, and w1, w2, w3 are the weight coefficients;

[0133] (9) the expected weighted sum S' and the expected weighted difference D' of each evaluation index are set;

[0134] (10) the change of each evaluation index is determined based on S, S', D and D'.

[0135] In practical applications, we can set these weights according to the user's preferences. For example, if the user cares more about battery life, we can set w3 to be larger.

[0136] The present application also proposes an adjustment strategy based on these indicators. For example, if the weighted sum S is less than the expected value S', we will increase the wireless transmission power and CPU frequency; if the weighted difference D is greater than the expected value D', we will also increase these two parameters; if D is less than D', we will keep the wireless transmission power unchanged and only slightly increase the CPU frequency. This strategy can find a good balance point between multiple performance indicators.

[0137] In another embodiment of the present application, the method of determining the power consumption management strategy in step (10) comprises:

[0138] if S < S', increase the wireless transmission power PTX and the CPU frequency FCPU;

[0139] if D > D', increase the wireless transmission power PTX and the CPU frequency FCPU;

[0140] if D < D', keep the wireless transmission power PTX at the current value and keep the CPU frequency FCPU higher than the current value.

[0141] In another embodiment of the present application, the method of determining the change of each evaluation index in step (10) comprises:

[0142] F CPU_new = F CPU

[0143] F CPU-new = F CPU + a1

[0144]

[0145] If the corresponding evaluation index is CPU frequency, then: If the corresponding evaluation index is wireless transmission power, then:

[0146] P TX_new = P TX + b

[0147]

[0148] If the corresponding evaluation index is audio processing complexity C Audio , then:

[0149]

[0150] where a max , b max and c max are the maximum allowed adjustment amplitudes of CPU frequency, wireless transmission power and audio processing complexity, respectively, η, ρ and ω are adjustment coefficients, F′ CPU , P′ TX and C′ Audio are the corresponding target values.

[0151] In another embodiment of the present application, the following steps are further included:

[0152] (11) Determine whether the following condition is satisfied:

[0153] max(x t , x t+1 , …, x t+T-1 ) > (1 + θ) * x t+T

[0154] where θ is a power fluctuation tolerance coefficient, θ > 0;

[0155] (12) If the condition is satisfied, perform smoothing processing:

[0156] x t+i_smooth = λx t+i + (1 - λ) x t+T

[0157] where i = 0, 1, …, T - 1, λ is a smoothing coefficient, 0 < λ < 1;

[0158] (13) Re-execute steps (2) to (5) using the smoothed power consumption data.

[0159] In another embodiment of the present application, the following steps are further included:

[0160] (14) periodically updating the RNN model parameters, the updating method being: using the actual power consumption data and the predicted power consumption data of the last N time windows to construct a training set;

[0161] The backpropagation algorithm is used to minimize the loss function L:

[0162]

[0163] where x i +T is the actual power consumption, is the predicted power consumption;

[0164] The weight matrix W h is updated as follows: hh x、W hy 、W h and the bias vector b y are updated as follows:

[0165]

[0166] where η is the learning rate, W represents any weight matrix, and b represents any bias vector.

[0167] This periodic updating method can make the RNN model continuously adapt to new power consumption patterns and maintain the accuracy of the prediction.

[0168] In summary, the Bluetooth headset neural network power consumption management method proposed in the present application has the following main advantages: first, it can accurately predict future power consumption, providing an important basis for power consumption management; second, it uses a multi-level judgment and adjustment strategy, which can respond appropriately to different power consumption situations; third, it considers multiple performance indicators, which can find a good balance point between power consumption, performance and user experience; finally, it has adaptive ability, which can continuously learn and adjust to adapt to different user habits. These characteristics make the present application have strong practicality and advanced nature in practical application.

[0169] Example 1: In one of the best embodiments of the present application, we selected a common Bluetooth headset model on the market as the test object. The headset uses a Qualcomm QCC5141 chip, supports Bluetooth 5.2 protocol, and has active noise reduction function. We used the neural network power consumption management method proposed in the present application to optimize the headset.

[0170] Specifically, we first collected historical power consumption data of the earphone under different usage scenarios, including music playback, call, and standby state. We chose 10 minutes as the prediction time window T, and recorded the power consumption data every 30 seconds. Then, we constructed a two-layer LSTM (Long Short-Term Memory) network as our RNN prediction model, with 64 hidden layer neurons. We used the Adam optimizer with a learning rate of 0.001 and pre-trained on the collected historical data.

[0171] In actual operation, we updated the RNN model parameters every 5 minutes to adapt to changes in user usage habits. For the power consumption adjustment strategy, we set μ = 0.9 as the long-term power consumption evaluation coefficient and α = 0.5 as the adjustment coefficient. The maximum adjustable amplitude amax of CPU frequency was set to 200 MHz, and the maximum adjustable amplitude bmax of wireless transmission power was set to 5 mW.

[0172] Comparative Example 1: As a comparison, we chose the same model of Bluetooth earphone but used the traditional fixed threshold power consumption management method. This method sets three fixed power consumption thresholds: low, medium, and high. When the power consumption is detected to be lower than the low threshold, the CPU frequency and wireless transmission power are increased; when the power consumption is higher than the high threshold, these two parameters are decreased; when the power consumption is in the middle range, the parameters remain unchanged.

[0173] Detection methods and standards: To comprehensively evaluate the performance of the method of the present invention, we designed the following detection indicators and methods:

[0174] 1. Average power consumption: The average power consumption of the earphone in 1 hour was measured using a high-precision power meter.

[0175] 2. Battery life: The time from full charge to automatic shutdown of the earphone was measured under the condition of a new battery.

[0176] 3. Audio delay: The delay time of audio signals from transmission to reception was measured using professional audio analysis equipment.

[0177] 4. User experience score: 50 volunteers were invited to use the earphones optimized by the two methods and score the sound quality, battery life, comfort, etc., with a full score of 10.

[0178] We conducted a one-week test, repeating the above tests in different usage scenarios every day. The following are the average values of the test results:

[0179] Indicator Method of the invention Conventional method Lift percentage (%) Average power consumption (mW) 15.2 18.7 18.7 Battery life (h) 8.5 7.2 18.1 Audio delay (ms) 112 135 17 User experience score 8.7 7.5 16

[0180] The test results are analyzed as follows:

[0181] From the above test results, we can see that the neural network power management method proposed in the present application is significantly better than the traditional fixed threshold method in various aspects.

[0182] First, in terms of average power consumption, the method of the present application reduces by 18.7% compared with the traditional method. This significant reduction is mainly due to the fact that our RNN prediction model can accurately predict future power consumption requirements, thereby avoiding unnecessary power consumption while ensuring performance. The traditional method, due to its passive response, often overestimates performance when power consumption is low, resulting in energy waste.

[0183] Secondly, the battery endurance time is improved by 18.1%, which directly reflects the actual benefits brought by the reduction of power consumption. For portable devices such as Bluetooth earphones, longer battery life means better user experience and higher product competitiveness.

[0184] In terms of audio delay, our method also achieved an improvement of 17.0%. This is because our method can appropriately increase CPU frequency and wireless transmission power when power consumption allows, thereby reducing the delay of audio processing and transmission. In contrast, the traditional method may excessively reduce these parameters in high power consumption situations, resulting in higher delay.

[0185] Finally, the user experience score is improved by 16.0%, which is a comprehensive indicator reflecting the advantages of our method in actual use. Lower power consumption, longer battery life, and lower delay, while the traditional method performs poorly in these aspects, thus affecting the overall user experience.

[0186] It is worth noting that our method has achieved more than 15% improvement in various indicators, which shows the comprehensiveness and balance of the method. It is not just an improvement in one aspect, but a significant progress in multiple key indicators. This comprehensive improvement is due to the core innovation of our method: using neural networks for accurate power consumption prediction and making multi-level and dynamic adjustments based on the prediction results.

[0187] In addition, although not directly reflected in the table, we observed during the test that the performance of the earphone is more stable after using our method, and it does not have the common high and low fluctuations as in the traditional method. This stability is also very important for user experience.

[0188] Overall, these test results fully demonstrate the superiority of the method of the present application. It not only surpasses the traditional method in technical indicators, but also shows obvious advantages in actual user experience. This intelligent and adaptive power management method points out a new direction for the development of Bluetooth earphones and other portable electronic devices.

[0189] It should be noted that the above only describes the preferred embodiments of the present application and is not intended to limit the present application, and any modification, equivalent replacement, improvement, etc. made within the principles of the present application shall be included in the protection scope of the present application.

Claims

1. A neural network power management method for Bluetooth headsets, characterized in that, It includes the following steps: (1) Obtain the historical power consumption dataset X = {x t ,x t+1 ,…,x t+T-1 }, where x t Let t be the power consumption at time t, and T be the prediction time window; (2) Predict the current power consumption x using a recurrent neural network (RNN). t+T The prediction model of RNN is as follows: h t =tanh(W hx x t +W hh h t-1 +b h ) y t =W hy h t +b y x t+T =σ(y t ) Among them, h t Let W be the hidden state at time t. hx W is the weight matrix input into the hidden layer. hh W is the weight matrix from hidden layer to hidden layer. hy Let b be the weight matrix from the hidden layer to the output layer. h and b u σ represents the bias vectors of the hidden layer and the output layer, respectively, and σ is the sigmoid activation function. (3) The predicted current power consumption x t +T and current power consumption requirements x t Compare with +T′; (4) Determine the adjustment strategy based on the comparison result; (5) Adjust the CPU frequency F of the Bluetooth headset according to the adjustment strategy. CPU and wireless transmission power P TX .

2. The neural network power consumption management method for Bluetooth headsets as described in claim 1, characterized in that, The step (4) of determining the adjustment strategy based on the comparison results includes: if x t+T <x′ t+T Then increase the wireless transmission power P TX and CPU frequency F CPU ; Otherwise, determine whether the following conditions are satisfied: (x t +x t+1 +…+x t+T-1 )>μx t+T where μ is the long-term power consumption evaluation coefficient, and 0 < μ < 1.

3. The neural network power consumption management method for Bluetooth headsets as described in claim 2, characterized in that, When (x) is satisfied t +x t+1 +…+x t+T-1 )>μx t+T The process also includes the following steps: (a) Calculate the average power consumption y over the time period t : Where y = 1, 2, ..., T-1; (b) The obtained T-1 average power consumption y t The power consumption demand y for the corresponding time period is respectively t Compare; (c) If y exists t <y′ t Then adjust the current power consumption, which is higher than the power consumption requirement, to the corresponding expected current power consumption x″. t+T and increase wireless transmission power P TX and CPU frequency F CPU ; (d) If y does not exist t <y′ t Then, based on the current power consumption x t+T Find x′ with the corresponding power consumption requirement. t+T Determine the adjustment strategy.

4. The neural network power consumption management method for Bluetooth headsets as described in claim 3, characterized in that, The method for adjusting the current power consumption in step (c) is as follows: x″ t+T =x′ t+T +α*(x t+T -x′ t+T ) where α is the adjustment coefficient, 0 < α < 1, and is used to control the adjustment amplitude.

5. The neural network power consumption management method for Bluetooth headsets as described in claim 1, characterized in that, The method for adjusting the CPU frequency FCPU and the wireless transmission power PTX in step (5) is as follows: F CPU_new =F CPU +a1 P TX_new =P TX +b1 where, a max and b max These represent the maximum adjustable ranges of CPU frequency and wireless transmission power, respectively, with β and γ being adjustment coefficients, and P... c and P w These are the power consumption coefficients for the CPU and the wireless transmission module, respectively.

6. The neural network power consumption management method for Bluetooth headsets as described in claim 1, characterized in that, It further includes the following steps: (6) Calculate the performance P, latency L, and battery life B as evaluation metrics; (7) Determine whether any of the evaluation metrics exceeds the corresponding preset threshold. If not, end; if so, execute step (8); (8) Determine the weighted sum S and weighted difference D of performance, latency, and battery life: S = w1P + w2L + w3B D = |w1P - w2L| + |w2L - w3B| + |w3B - w1P| where P, L, and B respectively represent the normalized values of performance, latency, and battery life, and w1, w2, and w3 are weight coefficients; (9) Set the expected weighted sum S for each evaluation indicator. ′ And the expected weighted difference D ′ (10) Based on S, S ′ D and D ′ Determine the amount of change corresponding to each evaluation indicator.

7. The neural network power consumption management method for Bluetooth headsets as described in claim 6, characterized in that, The method for determining the power management strategy in step (10) includes: If S < S', increase the wireless transmission power PTX and the CPU frequency FCPU; If D > D', increase the wireless transmission power PTX and the CPU frequency FCPU; If D < D', keep the wireless transmission power PTX at the current value and keep the CPU frequency FCPU higher than the current value.

8. The neural network power consumption management method for Bluetooth headsets as described in claim 6, characterized in that, The method for determining the change amount corresponding to each evaluation metric in step (10) includes: F CPU-new =F CPU F CPU-new =F CPU +a1 If the corresponding evaluation metric is the CPU frequency, then: If the corresponding evaluation metric is the wireless transmission power, then: P TX_new =P TX +b1 If the corresponding evaluation metric is audio processing complexity C Audio ,but: Where a max b max and c max These represent the maximum allowable adjustment ranges for CPU frequency, wireless transmission power, and audio processing complexity, respectively, where η, ρ, and ω are adjustment coefficients, and F′ is the maximum allowable adjustment range for these parameters. CPU 、P′ TX and C′ Audio This is the corresponding target value.

9. The neural network power consumption management method for Bluetooth headsets as described in claim 1, characterized in that, It further includes the following steps: (11) Determine whether the following conditions are satisfied: max(x t ,x t+1 ,…,x t+T-1 )>(1+θ)*x t+T where θ is the power consumption fluctuation tolerance coefficient, θ > 0; (12) If the conditions are satisfied, perform smoothing processing: x t+i_smooth =λx t+i +(1-λ)x t+T where i = 0, 1,..., T - 1, λ is the smoothing coefficient, 0 < λ < 1; (13) Use the smoothed power consumption data to re-execute steps (2) to (5).

10. The neural network power consumption management method for Bluetooth headsets as described in claim 1, characterized in that, It further includes the following steps: (14) Regularly update the RNN model parameters, and the update method is as follows: Construct a training set using the actual power consumption data and predicted power consumption data of the most recent N time windows; Adopt the backpropagation algorithm to minimize the loss function L: Where x i +T represents the actual power consumption. To predict power consumption; Update weight matrix W h x、W hh W hy and bias vector b h ,b y : where η is the learning rate, W represents any weight matrix, and b represents any bias vector.

Images